memstead_cli/commands/projection.rs
1//! `memstead projection` — the binding (projection-promotion) command tree.
2//!
3//! The projection is the unit: one versioned binding per source→mem obligation
4//! (bundle plan `03-projection-promotion`). The tree ships nine leaves —
5//! `brief`, `init`, `migrate`, `enable`, `edit`, `advance`, `exclude`,
6//! `verify`, `check-path`:
7//!
8//! - `brief` renders a binding's run-brief — the Markdown prompt an agent
9//! consumes — for a canonical binding id `<mem>/<stem>` (D3/D9), or the next
10//! due binding under `--all` (round-robin + backoff selection).
11//! - `init` scaffolds a fresh v2 single-record binding non-interactively.
12//! - `migrate` converts every prior on-disk generation into v2 records in
13//! place: gen-1 root folders, the gen-2 four-primitive store, and the v1
14//! three-file store — folding medium+facet content inline.
15//! - `enable` adds a missing `build` / `sync` / `verify` operation block to an
16//! existing binding (D6 — the remedy a refused mutating op cites).
17//! - `edit` patches an existing binding's author-editable fields in place.
18//! - `advance` records disposition-gated sync-baseline advances (D7).
19//! - `exclude` records authored exclusions for in-scope source artifacts.
20//! - `verify` measures a binding's fidelity, records durable findings, and
21//! renders the tier-1 fidelity report (E3b).
22//! - `check-path` answers deny verdicts: is a path or pattern hidden by a
23//! binding's `deny_paths`?
24//!
25//! This tree is the sole binding surface: the retired `ingest` and `pipeline`
26//! command trees folded in here (`ingest brief` → `projection brief`,
27//! `pipeline migrate` → `projection migrate`'s gen-1 path).
28//!
29//! Errors carry `PROJECTION_*` wire tokens (D12); the missing-workspace path is
30//! single-sourced through [`crate::setup::workspace_not_initialised_error`].
31
32use clap::{Args as ClapArgs, Subcommand, ValueEnum};
33use serde_json::json;
34
35use memstead_base::binding::{
36 BuildMode, BuildOperation, CapabilityError, DEFAULT_ADJUDICATION_CAP,
37 DEFAULT_FULL_RESYNC_EVERY, ScaffoldParams, SyncOperation, VerifyOperation, validate_binding,
38};
39use memstead_base::binding_migrate::{
40 BindingMigrateError, check_all_consumed, fold_v1_binding, migrate_gen2_bindings,
41};
42use memstead_base::ingest::advance::{
43 AdvanceError, DispositionInput, ExcludeError, advance_baseline, record_exclusions,
44};
45use memstead_base::ingest::findings::{
46 FindingsError, FullResyncDecision, record_anchor_hash_backfill, record_verified_baseline,
47 verify_binding, verify_binding_full,
48};
49use memstead_base::ingest::report::{
50 DEFAULT_REPORT_BUDGET, compute_fidelity_report, render_fidelity_report,
51};
52use memstead_base::ingest::resolve::{ResolveError, ResolvedSource, resolve_binding_run};
53use memstead_base::ingest::{
54 OperationFilter, OperationKind, RenderBriefError, not_loop_declared, render_ingest_brief,
55 render_sync_brief_for, render_verify_brief_for, select_next_due_operation,
56};
57use memstead_base::pipeline::{IngestTrigger, MediumType};
58use memstead_base::pipeline_store::{
59 ProjectionGeneration, delete_ingest, load_legacy_pipeline_configs, load_pipeline_configs,
60 load_projection_generations, read_binding, remove_mediums_and_facets_trees, write_binding,
61};
62use memstead_base::workspace_store::StoreError;
63use memstead_base::{migrate_legacy_pipeline, read_legacy_pipeline_configs};
64
65use crate::CliError;
66use crate::output::{ExitKind, print_json, print_markdown};
67
68/// Version marker on `projection verify --json`. External contract: bump the
69/// `vN` when the payload's shape changes so a consumer asserting it fails
70/// loudly instead of misparsing. House style — see `memstead-export/v1`
71/// (`commands/export.rs`) and `workspace-dump/v1` (`commands/workspace.rs`).
72pub const JSON_VERIFY_FORMAT: &str = "memstead-verify/v1";
73use crate::setup::{CliContext, workspace_not_initialised_error};
74
75#[derive(ClapArgs, Debug)]
76pub struct Args {
77 #[command(subcommand)]
78 pub command: ProjectionCommand,
79}
80
81#[derive(Subcommand, Debug)]
82pub enum ProjectionCommand {
83 /// Render a binding's run-brief — the Markdown prompt an agent consumes —
84 /// on stdout. Takes the canonical binding id `<mem>/<stem>` (D3), e.g.
85 /// `engine/graph`. Omit the id (or pass `--all`) to select the next due
86 /// (binding, operation) pair by round-robin + backoff and render that
87 /// operation's brief; `--operation` picks which operations rotate (default
88 /// `build` — the classic build-only rotation; `any` rotates every
89 /// loop-declared build / sync / verify pair). An operation participates
90 /// only where its binding block declares `trigger: loop`. Reads the v2
91 /// binding store and the destination mem's schema / writing guidance; the
92 /// assembly lives in the engine, so every consuming surface renders
93 /// byte-identical briefs by construction.
94 ///
95 /// `--verify` renders the **verify brief** (group C) for the named binding:
96 /// measurement + capped-adjudication instructions only, with no
97 /// destination-mutation instruction. `--sync` renders the **sync brief** —
98 /// the sole maintenance-writer prompt, carrying both the cursor slice and the
99 /// open verify findings in one brief with the absorbed reconcile
100 /// conservatism. Both are read-only on the mem; the sync brief's repairs
101 /// reach the mem only when an agent acts on it through the MCP mutation
102 /// surface.
103 Brief(BriefArgs),
104 /// Scaffold a fresh v2 binding non-interactively: ONE record with one
105 /// inline source, at `.memstead/projections/<mem>/<stem>.json`.
106 /// All inputs are flags — no prompts ever (parity across callers). The
107 /// default binding declares build+sync+verify where the medium permits:
108 /// a `web` source scaffolds build-only, with the deferral named in
109 /// `warnings[]`. A `prune` block is scaffolded wherever sync survived,
110 /// with the strongest guarantee the medium supports (never-clobber for a
111 /// git-backed source). Refuses `PROJECTION_EXISTS` (without touching disk)
112 /// when a binding of the same id already exists — never overwrites.
113 Init(InitArgs),
114 /// Migrate every prior on-disk generation into v2 single-record
115 /// bindings, in place. Gen-1 — the root-folder
116 /// `scopes|projections|ingests/` JSON layout — is first materialized
117 /// into the four-primitive store, then folded. Gen-2 — the
118 /// four-primitive store (per-mem `Projection` + flat `Ingest`) — merges
119 /// each ingest into its projection and folds the referenced facets +
120 /// mediums inline. v1 — the three-file store — folds each binding's
121 /// facet references inline the same way, source names preserved
122 /// byte-verbatim (they key sync watermarks). The emptied `mediums/` and
123 /// `facets/` trees are removed; orphan records refuse rather than drop.
124 /// `refinement` mode and dangling refs refuse with a typed error.
125 /// Idempotent on a migrated store. Use `--dry-run` to preview without
126 /// writing.
127 Migrate(MigrateArgs),
128 /// Enable a `build` / `sync` / `verify` operation on an existing binding by
129 /// adding its block (with sensible defaults) if absent. This is the remedy
130 /// a refused *mutating* operation cites (D6): `projection enable sync
131 /// <binding>`. Before writing, the operation is checked against the
132 /// medium-capability matrix (D6) — enabling `sync`/`verify` over a medium
133 /// that cannot support it (e.g. a `web` source) refuses with the capability
134 /// gap and writes nothing. Enabling an already-present operation refuses
135 /// `PROJECTION_OP_ALREADY_ENABLED`; a missing binding refuses
136 /// `PROJECTION_NOT_FOUND`.
137 Enable(EnableArgs),
138 /// Patch an existing binding's author-editable fields in place — the
139 /// general edit surface over the shared `pipeline_edit` layer that
140 /// `init`/`enable` already use. The patch is a JSON object with patch
141 /// semantics: an absent field is preserved, an explicit `null` clears
142 /// `intent`/`rules`/`prune`, and a present `sources`, `operations`,
143 /// `reference_mems` or `deny_paths` value replaces that whole block
144 /// (`version` stays engine-managed and is ignored if supplied). Every
145 /// edit is validated before anything lands: a patch that would
146 /// introduce a refusal the stored record does not already carry (a
147 /// duplicate source name, `sync` over a medium with no change signal)
148 /// refuses with the capability gap and writes nothing — pre-existing
149 /// refusals never block an unrelated edit. Refuses
150 /// `PROJECTION_NOT_FOUND` for a missing binding; adding a source to a
151 /// binding is `edit <binding> --patch '{"sources":[...]}'` with the
152 /// FULL source list, existing entries included, since the block
153 /// replaces.
154 Edit(EditArgs),
155 /// Advance a binding's sync baseline by recording per-artifact
156 /// dispositions (D7). The engine freezes the presented changed slice,
157 /// subtracts already-disposed artifacts on re-presentation, appends
158 /// new-HEAD deltas when the source moves mid-pass, and — when the
159 /// remainder empties — advances the destination mem's `#synced` token via
160 /// the sync-state writer (provenance piggybacks that commit). Dispositions
161 /// are durable (`.memstead/state/advance/`), so a partial pass resumes
162 /// across process restarts. The gate accepts **only** artifact ids the
163 /// engine presented — an unknown id refuses the whole call atomically
164 /// (`PROJECTION_ADVANCE_UNKNOWN_ARTIFACT`). In this cycle the agent supplies
165 /// a disposition for **every** artifact explicitly (auto-derivation lands
166 /// later).
167 Advance(AdvanceArgs),
168 /// Declare authored **exclusions** for in-scope source artifacts. Unlike
169 /// `advance` (whose gate accepts only artifacts in the changed slice), this
170 /// gates on enumerable `S(D)` membership, so a stable, unchanged artifact can
171 /// be recorded as deliberately not-modeled with a rationale. Each accepted
172 /// `(artifact, rationale)` lands in the durable exclusion ledger the fidelity
173 /// report consults, so the artifact stops re-surfacing as `uncovered` under
174 /// exhaustive coverage and keeps its reasoning. An id may be written in
175 /// either form, the workspace-relative one the report uses or the
176 /// source-relative one (relative to the source's pointer): it is resolved
177 /// through the binding's source join at exclude time, the ledger holds the
178 /// canonical workspace-relative id, and the response lists what was
179 /// recorded. An id that resolves to no artifact of `S(D)` refuses the
180 /// whole call atomically (`PROJECTION_EXCLUDE_NOT_SOURCE_MEMBER`, naming
181 /// the nearest known ids); re-declaring merges into the ledger. On a
182 /// binding with several primary sources, a source-relative id that
183 /// resolves under MORE THAN ONE of them is likewise refused whole
184 /// (`PROJECTION_EXCLUDE_AMBIGUOUS_ARTIFACT`), naming every canonical id
185 /// it could denote in the message and under `details.ambiguous`: the
186 /// engine never picks a source the caller did not name, and either
187 /// canonical id is the unambiguous recovery. The write
188 /// path for the option-(a)
189 /// process-mem judgment migration, and the general "this in-scope artifact is
190 /// mined and warrants no destination entity, because …" capability.
191 Exclude(ExcludeArgs),
192 /// Measure a binding's fidelity and record durable findings (E3b, group A).
193 /// Read-only on the destination mem's ENTITIES: verify adjudicates its anchors
194 /// against the live source and samples in-scope artifacts, writing findings
195 /// keyed by the binding's `hash(D)` alone into the engine-owned findings store
196 /// (`.memstead/state/findings/`); the `source_head` a finding was observed
197 /// at is carried as metadata, so findings survive source-head movement and
198 /// stay presented until a later verify supersedes them. A
199 /// binding-declaration edit changes `hash(D)` and segregates prior
200 /// findings as superseded, never presented as current. Verify mutates no entity — any
201 /// repair routes through the (later) sync brief — but it is not a pure read:
202 /// a completed run records the findings store and backfills observed content
203 /// hashes onto hash-less anchors (measurement machinery, without which an anchor
204 /// never leaves `recheck`). The `#verified` freshness baseline moves only under
205 /// `--advance`, so a gate that verifies in order to read leaves the mem's config
206 /// byte-identical. It then renders the
207 /// deterministic, token-budgeted **tier-1 fidelity report** (group B) over
208 /// the findings just recorded: grain-classed coverage (an artifact ruled out
209 /// by `projection exclude` is dropped from `coverage.uncovered` and counted
210 /// beside it as `coverage.excluded`, so the uncovered figure is the number
211 /// still owed) with tree-anchor
212 /// fan-out on its own axis, anchor-resolution %, freshness vs. both
213 /// `sync_state` tokens (`signal: none` → freshness unknowable), the
214 /// capability-matrix block, and the tier-3 backlog depth — aggregates always
215 /// ship; heavy per-artifact lists greedy-fill under `--budget` and drop to
216 /// hints (forced back in with `--include`).
217 ///
218 /// The anchor figures answer for THIS BINDING'S population only: anchors
219 /// another binding wrote, and anchors pointing outside this binding's
220 /// declared scope, are excluded from every figure and named in the report
221 /// rather than dropped. They are never deleted or rewritten: exclusion is
222 /// a reporting decision. The report states what its denominator counted
223 /// (anchor rows, with the distinct-artifact count beside them, since one
224 /// artifact legitimately carries several rows), and says so when anchors
225 /// recording no producing binding are included by the pre-provenance
226 /// fallback.
227 ///
228 /// A destination mem that is QUARANTINED (its schema pin unresolved, its
229 /// mount unbacked) refuses the run outright with `MEM_QUARANTINED` and its
230 /// boot reason, on both `verify` and `brief --verify`: a mem serving no
231 /// entities would read as every artifact uncovered and every anchor
232 /// absent, which is a measurement neither can honestly make. The refusal
233 /// is placed above every store the run would otherwise write, so the
234 /// findings store, the `#verified` token and the authored exclusions in
235 /// the advance file all come through byte-identical.
236 ///
237 /// A sidecar row whose ENTITY the mem no longer holds is reported as
238 /// dangling and named, in no figure and never as resolving: it is a
239 /// sidecar integrity condition, not an anchor state, and nothing repairs
240 /// it, because the row is the trace of a writer that went around the
241 /// engine. Where the entity end could not be reconciled at all (an
242 /// unloaded, quarantined or partly unparsed mem), the report says so
243 /// rather than reporting a clean anchor axis.
244 Verify(VerifyArgs),
245 /// Answer deny verdicts: is a path (or Glob/Grep pattern) hidden by a
246 /// binding's `deny_paths`? Evaluates each candidate against the named
247 /// binding — or, with `--binding` omitted, the ACTIVE binding (the one
248 /// whose brief was last consumed) — using the engine's own deny dialect:
249 /// the facet-scope glob grammar, resolved against the workspace root (an ingest deny spans every source, so it has no pointer to be relative to), plus
250 /// the literal-base directory-prefix rule (`dev/**` also blocks a read of
251 /// `dev` itself). Single-path form takes the candidate as an argument;
252 /// `--batch` reads `{"cwd": "<dir>", "paths": ["...", ...]}` from stdin
253 /// and answers every candidate in one process — the form a per-tool-call
254 /// consumer (the plugin's PreToolUse deny hook) amortizes subprocess cost
255 /// with. Engine-free and read-only: verdicts come from the binding record
256 /// and the path alone, no workspace boot. Output names the matched deny
257 /// entry on a block; the exit code stays 0 for an answered check — a
258 /// non-zero exit means the check itself could not run (unknown or
259 /// quarantined binding, no active binding, malformed batch).
260 CheckPath(CheckPathArgs),
261}
262
263/// The medium type flag for `projection init` — the CLI-facing mirror of
264/// [`MediumType`] (which carries serde, not clap, derives). Decides the
265/// capability matrix (D6) that filters the default binding's operations.
266#[derive(Clone, Copy, Debug, ValueEnum)]
267pub enum MediumTypeArg {
268 /// A source tree of code.
269 Codebase,
270 /// A directory of files (non-code).
271 Filesystem,
272 /// A git history.
273 Git,
274 /// Another mem's graph.
275 Graph,
276 /// Web sources (build-only this cycle — no change signal).
277 Web,
278}
279
280impl MediumTypeArg {
281 fn to_medium_type(self) -> MediumType {
282 match self {
283 MediumTypeArg::Codebase => MediumType::Codebase,
284 MediumTypeArg::Filesystem => MediumType::Filesystem,
285 MediumTypeArg::Git => MediumType::Git,
286 MediumTypeArg::Graph => MediumType::Graph,
287 MediumTypeArg::Web => MediumType::Web,
288 }
289 }
290}
291
292#[derive(ClapArgs, Debug)]
293pub struct BriefArgs {
294 /// The canonical binding id `<mem>/<stem>` (D3) — e.g. `engine/graph`.
295 /// Omit (or pass `--all`) to select the next due binding by round-robin +
296 /// backoff. Required with `--verify` / `--sync` (those operate on one
297 /// binding's live findings/cursor, never a rotation).
298 pub binding: Option<String>,
299 /// Select the next due (binding, operation) pair across all bindings
300 /// (round-robin + backoff) and render its brief, instead of naming one.
301 /// Which operations rotate is decided by `--operation` (default: build
302 /// only). Ignored with `--verify` / `--sync`.
303 #[arg(long)]
304 pub all: bool,
305 /// Which operations the `--all` rotation considers. An operation
306 /// participates only where the binding declares its block with
307 /// `trigger: loop` — consent lives in the declaration. `build` (the
308 /// default) keeps the classic build-only rotation; `any` rotates across
309 /// every loop-declared build / sync / verify pair and renders the matching
310 /// brief (the `--json` output names the picked operation).
311 #[arg(long, value_enum, default_value_t = BriefOperationArg::Build, requires = "all", conflicts_with_all = ["verify", "sync"])]
312 pub operation: BriefOperationArg,
313 /// Take the rotation slot this `--all` render selects — advancing the
314 /// round-robin cursor and the per-pair backoff state. WITHOUT this flag a
315 /// `--all` render is a pure read: repeat it as often as you like and the
316 /// rotation stays exactly where it was (a diagnostic re-render must never
317 /// silently skip the next binding). The loop driver that will act on the
318 /// brief passes `--consume`; nothing else should.
319 #[arg(long, requires = "all")]
320 pub consume: bool,
321 /// Render the **verify brief** (group C) for the named binding instead of
322 /// the build brief: measurement + capped-adjudication instructions only.
323 /// It carries no destination-mutation instruction — repairs route through
324 /// the sync brief. Read-only on the mem. Mutually exclusive with `--sync`.
325 #[arg(long, conflicts_with = "sync")]
326 pub verify: bool,
327 /// Render the **sync brief** (group C) for the named binding instead of the
328 /// build brief: the sole maintenance-writer prompt, carrying both the cursor
329 /// slice and the open verify findings in one brief, with the absorbed
330 /// reconcile conservatism. Read-only on the mem (the agent's writes route
331 /// through MCP). Mutually exclusive with `--verify`.
332 #[arg(long, conflicts_with = "verify")]
333 pub sync: bool,
334}
335
336/// The `--operation` value for `projection brief --all` — which operations the
337/// rotation considers. CLI-facing mirror of the engine's [`OperationFilter`]
338/// (which carries no clap derives).
339#[derive(Clone, Copy, Debug, ValueEnum, PartialEq, Eq)]
340pub enum BriefOperationArg {
341 /// Rotate over build pairs only (the default — the classic rotation).
342 Build,
343 /// Rotate over sync pairs only.
344 Sync,
345 /// Rotate over verify pairs only.
346 Verify,
347 /// Rotate over every loop-declared build / sync / verify pair.
348 Any,
349}
350
351impl BriefOperationArg {
352 fn to_filter(self) -> OperationFilter {
353 match self {
354 BriefOperationArg::Build => OperationFilter::Only(OperationKind::Build),
355 BriefOperationArg::Sync => OperationFilter::Only(OperationKind::Sync),
356 BriefOperationArg::Verify => OperationFilter::Only(OperationKind::Verify),
357 BriefOperationArg::Any => OperationFilter::Any,
358 }
359 }
360}
361
362#[derive(ClapArgs, Debug)]
363pub struct InitArgs {
364 /// Destination mem the binding writes into — the `<mem>` half of the
365 /// binding id `<mem>/<stem>` and the per-mem tier the three files live under.
366 #[arg(long)]
367 pub mem: String,
368 /// The medium pointer — a path (codebase / filesystem / git) or a mem id /
369 /// URL (graph / web). Becomes the scaffolded medium's `pointer`.
370 #[arg(long)]
371 pub source: String,
372 /// The medium type — decides the capability matrix (D6) that filters which
373 /// operations the default binding declares.
374 #[arg(long = "medium-type", value_enum)]
375 pub medium_type: MediumTypeArg,
376 /// Intent prose for the agent (the binding's `intent`). Optional.
377 #[arg(long)]
378 pub intent: Option<String>,
379 /// Binding stem — the `<stem>` half of the binding id and the file name
380 /// of the scaffolded binding record. Defaults to the final
381 /// path component of `--source`.
382 #[arg(long)]
383 pub name: Option<String>,
384}
385
386#[derive(ClapArgs, Debug)]
387pub struct MigrateArgs {
388 /// Preview the produced bindings (and any warnings) without writing them
389 /// to disk or removing the merged ingest files.
390 #[arg(long)]
391 pub dry_run: bool,
392}
393
394/// The operation `projection enable` adds to a binding. Mirror of the binding's
395/// operations block: `build` is always present (required), so enabling it
396/// always refuses as already-enabled; `sync` / `verify` are the enableable
397/// blocks.
398#[derive(Clone, Copy, Debug, ValueEnum, PartialEq, Eq)]
399pub enum EnableOperationArg {
400 /// The build operation (always present — enabling refuses as already-enabled).
401 Build,
402 /// The sync (maintenance-write) operation.
403 Sync,
404 /// The verify (measurement) operation.
405 Verify,
406}
407
408impl EnableOperationArg {
409 fn name(self) -> &'static str {
410 match self {
411 EnableOperationArg::Build => "build",
412 EnableOperationArg::Sync => "sync",
413 EnableOperationArg::Verify => "verify",
414 }
415 }
416}
417
418#[derive(ClapArgs, Debug)]
419pub struct EnableArgs {
420 /// The operation to enable: `build` | `sync` | `verify`.
421 #[arg(value_enum)]
422 pub operation: EnableOperationArg,
423 /// The binding id `<mem>/<stem>` (D3) — e.g. `engine/graph`.
424 pub binding: String,
425}
426
427#[derive(ClapArgs, Debug)]
428pub struct EditArgs {
429 /// The binding id `<mem>/<stem>` (D3) — e.g. `engine/graph`.
430 pub binding: String,
431 /// The JSON patch over the author-editable record, e.g.
432 /// `'{"sources": [ ...the full replacement list... ]}'`. Patch
433 /// semantics: absent = preserved, `null` clears where clearing is
434 /// legal, a present block replaces that block whole.
435 #[arg(long)]
436 pub patch: String,
437}
438
439#[derive(ClapArgs, Debug)]
440pub struct AdvanceArgs {
441 /// The binding id `<mem>/<stem>` (D3) — e.g. `engine/graph`.
442 pub binding: String,
443 /// A JSON object mapping each judged artifact id to its disposition, e.g.
444 /// `'{"src/lib.rs": "worked", "src/old.rs": "irrelevant"}'`. A value may
445 /// instead be an object carrying an authored rationale —
446 /// `'{"src/gen.rs": {"disposition": "excluded", "rationale": "generated, no entity"}}'`
447 /// — and an `excluded` verdict with a rationale is retained in the durable
448 /// exclusion ledger so the artifact stops re-surfacing as `uncovered` and
449 /// keeps its reasoning. Only ids the engine presented in the brief's changed
450 /// slice are accepted — an unknown id refuses the whole call. Omitting the
451 /// flag (or passing `'{}'`) records nothing and re-presents the remainder —
452 /// the shape of a first advance over an empty presented slice, which
453 /// exists only to write the baseline.
454 #[arg(long, default_value = "{}")]
455 pub dispositions: String,
456}
457
458#[derive(ClapArgs, Debug)]
459pub struct ExcludeArgs {
460 /// The binding id `<mem>/<stem>` (D3) — e.g. `project/graph`.
461 pub binding: String,
462 /// A JSON object mapping each in-scope source artifact id to the authored
463 /// rationale for excluding it, e.g.
464 /// `'{"docs/legacy.md": "superseded; no entity", "vendor/x.rs": "generated"}'`.
465 /// Every id must resolve to a member of the binding's enumerable source
466 /// `S(D)`, in the workspace-relative form or the source-relative one; the
467 /// ledger holds the canonical workspace-relative id. An id resolving to no
468 /// artifact refuses the whole call, naming the nearest known ids.
469 #[arg(long)]
470 pub exclusions: String,
471}
472
473#[derive(ClapArgs, Debug)]
474pub struct CheckPathArgs {
475 /// The candidate — a path (absolute, or relative to `--cwd`) or a
476 /// Glob/Grep pattern. Omitted in `--batch` mode.
477 #[arg(required_unless_present = "batch", conflicts_with = "batch")]
478 pub path: Option<String>,
479 /// Binding to check against, canonical `<mem>/<stem>`. Defaults to the
480 /// active binding — the one whose brief was last consumed; refuses typed
481 /// (`NO_ACTIVE_BINDING`) when none is.
482 #[arg(long)]
483 pub binding: Option<String>,
484 /// Batch mode: read one JSON object from stdin —
485 /// `{"cwd": "<dir>", "paths": ["...", ...]}` — and answer every
486 /// candidate. A malformed payload refuses whole (`INVALID_INPUT`), never
487 /// part-answers. The payload's `cwd` wins over `--cwd`.
488 #[arg(long)]
489 pub batch: bool,
490 /// Directory relative candidates resolve against (default: the process
491 /// working directory).
492 #[arg(long)]
493 pub cwd: Option<std::path::PathBuf>,
494}
495
496#[derive(ClapArgs, Debug)]
497pub struct VerifyArgs {
498 /// The binding id `<mem>/<stem>` (D3) — e.g. `engine/graph`.
499 pub binding: String,
500 /// Token budget for the tier-1 fidelity report's **heavy** content
501 /// (per-artifact lists). Aggregated counts always ship in addition; heavy
502 /// lists greedy-fill and drop to `## Hints` when they do not fit. Defaults
503 /// to the house envelope budget.
504 #[arg(long)]
505 pub budget: Option<usize>,
506 /// Force a heavy report section in past the budget (repeatable):
507 /// `uncovered_artifacts` | `tree_fanout` | `superseded_findings`.
508 #[arg(long = "include")]
509 pub include: Vec<String>,
510 /// Full measurement: walk the entire enumerable source `S(D)` (the
511 /// rotating sample scheduler is bypassed), treat the per-run adjudication
512 /// cap as unlimited, and perform the prepared-hash backfill — the
513 /// report's coverage and accuracy figures are computed over everything,
514 /// with no sampling or truncation caveat. Refuses (typed) rather than
515 /// render a fabricated-complete report in two cases, both per facet: a
516 /// facet whose medium is non-enumerable, and a facet whose medium claims
517 /// enumerability but whose own walk yields no artifacts — a sibling facet
518 /// that walked does not excuse it. Without this flag the capped/sampled
519 /// loop economics are unchanged.
520 #[arg(long)]
521 pub full: bool,
522 /// CI-gate mode: exit 6 when the completed run recorded findings, so a
523 /// pull-request gate can branch on three outcomes without parsing output
524 /// — 0 completed-and-clean, 6 completed-with-findings, any other nonzero
525 /// the measurement itself failed. The full report is rendered first,
526 /// either way. Opt-in by design: without this flag the exit behaviour is
527 /// byte-for-byte what it has always been, so no existing consumer breaks.
528 #[arg(long)]
529 pub fail_on_findings: bool,
530 /// CI-gate mode for blind runs: exit 6 (typed
531 /// `PROJECTION_VERIFY_INCONCLUSIVE`) when the completed run's rollup
532 /// verdict is `inconclusive` — no readable change signal, an empty
533 /// enumerated scope — so a gate cannot go green on a measurement that
534 /// never happened. Replaces the documented two-step verdict read for
535 /// opted-in callers. Evaluated after `--fail-on-findings` (a findings
536 /// run with both flags exits with the findings code). The full report
537 /// is rendered first. Opt-in by design: without this flag an
538 /// inconclusive run keeps its long-standing exit 0.
539 #[arg(long)]
540 pub fail_on_inconclusive: bool,
541 /// Advance the `<binding>/<facet>#verified` baseline, the freshness
542 /// token the selection loop reads to decide when a verify is due
543 /// again. Off by default: a bare verify measures without moving the
544 /// mem's config, so a gate or a grader that verifies in order to READ
545 /// leaves that config byte-identical and the next run still sees the
546 /// binding as unverified. Two writes deliberately do NOT ride this
547 /// flag, because neither is a freshness claim: the findings store,
548 /// which is the verify surface's own state outside the mem, and the
549 /// prepared-hash backfill onto hash-less anchors, which is measurement
550 /// machinery. Withhold that backfill and an anchor never leaves
551 /// `recheck`, so drift stops being adjudicated at all.
552 #[arg(long)]
553 pub advance: bool,
554}
555
556pub fn run(ctx: &CliContext, args: Args) -> anyhow::Result<()> {
557 match args.command {
558 ProjectionCommand::Brief(a) => brief(ctx, a),
559 ProjectionCommand::Init(a) => init(ctx, a),
560 ProjectionCommand::Migrate(a) => migrate(ctx, a),
561 ProjectionCommand::Enable(a) => enable(ctx, a),
562 ProjectionCommand::Edit(a) => edit(ctx, a),
563 ProjectionCommand::Advance(a) => advance(ctx, a),
564 ProjectionCommand::Exclude(a) => exclude(ctx, a),
565 ProjectionCommand::Verify(a) => verify(ctx, a),
566 ProjectionCommand::CheckPath(a) => check_path(ctx, a),
567 }
568}
569
570/// The capability gap that would stop `projection enable <operation>`, if
571/// there is one.
572///
573/// Asks the question `enable` asks: validate a CANDIDATE record carrying the
574/// operation. Validating the record as it stands answers nothing — the
575/// operation is absent, so there is no declaration for the matrix to refuse,
576/// and the gap stays invisible exactly where a caller needs it.
577fn operation_capability_gap(
578 binding: &memstead_base::binding::Binding,
579 operation: &'static str,
580) -> Option<String> {
581 let mut candidate = binding.clone();
582 let batch_size = candidate
583 .operations
584 .build
585 .as_ref()
586 .map_or(20, |b| b.batch_size);
587 match operation {
588 "sync" => {
589 candidate.operations.sync = Some(SyncOperation {
590 trigger: IngestTrigger::Manual,
591 batch_size,
592 });
593 }
594 "verify" => {
595 candidate.operations.verify = Some(VerifyOperation {
596 trigger: IngestTrigger::Manual,
597 batch_size,
598 adjudication_cap: DEFAULT_ADJUDICATION_CAP,
599 full_resync_every: DEFAULT_FULL_RESYNC_EVERY,
600 });
601 }
602 // `build` is carried by every medium — no capability row refuses it,
603 // so the enable remedy is always honest for it.
604 _ => return None,
605 }
606 validate_binding(&candidate).err().and_then(|refusals| {
607 refusals
608 .iter()
609 .find(|r| {
610 matches!(
611 r,
612 CapabilityError::OperationOutOfScope { operation: op, .. } if *op == operation
613 )
614 })
615 .map(|gap| gap.to_string())
616 })
617}
618
619/// The refusal for a binding that declares no `sync` operation.
620///
621/// `projection enable sync <binding>` is the remedy — but only where the
622/// medium *can* carry sync. Over a medium whose capability row refuses it (a
623/// `web` source), `enable` refuses too, so naming it bounces the reader:
624/// the refusal points at `enable`, `enable` points at the capability gap,
625/// and nothing they can do closes it. Where the gap exists it IS the answer,
626/// so it is what this says.
627///
628/// Both codes are spelled as literals here, at their construction sites, so
629/// the generated error index (xtask) keeps finding them.
630fn absent_sync_error(binding_id: &str, binding: &memstead_base::binding::Binding) -> CliError {
631 if let Some(gap) = operation_capability_gap(binding, "sync") {
632 return CliError::new(
633 ExitKind::Validation,
634 "PROJECTION_CAPABILITY_UNSUPPORTED",
635 format!(
636 "binding `{binding_id}` has no sync operation, and this medium \
637 cannot carry one: {gap}"
638 ),
639 )
640 .with_details(json!({ "binding": binding_id, "operation": "sync" }));
641 }
642 CliError::new(
643 ExitKind::Validation,
644 "PROJECTION_SYNC_NOT_ENABLED",
645 format!(
646 "binding `{binding_id}` has no sync operation — enable it with \
647 `memstead projection enable sync {binding_id}`"
648 ),
649 )
650 .with_details(json!({
651 "binding": binding_id,
652 "remedy": { "cli": format!("memstead projection enable sync {binding_id}") },
653 }))
654}
655
656/// Map a [`RenderBriefError`] to a typed CLI error (D12). Not-found bindings /
657/// facets / mediums exit `NotFound`; a malformed id is a `Validation` name
658/// error; config-load and mode-unsupported failures are generic. Codes are
659/// spelled as literals at each construction site so the generated error index
660/// (xtask) picks them up.
661fn map_brief_err(binding_id: &str, err: RenderBriefError) -> CliError {
662 let message = err.to_string();
663 let mapped = match &err {
664 RenderBriefError::ConfigLoad(_) => {
665 CliError::new(ExitKind::Generic, "PROJECTION_LOAD_FAILED", message)
666 }
667 // D6/AC4: the binding declares no build op — refuse with the
668 // `projection enable build` remedy the error message already carries.
669 RenderBriefError::BuildOperationAbsent { .. } => CliError::new(
670 ExitKind::Validation,
671 "PROJECTION_BUILD_NOT_ENABLED",
672 message,
673 ),
674 // A malformed findings store while rendering a verify / sync brief.
675 RenderBriefError::FindingsRead { .. } => CliError::new(
676 ExitKind::Generic,
677 "PROJECTION_FINDINGS_READ_FAILED",
678 message,
679 ),
680 RenderBriefError::Resolve(inner) => match inner {
681 ResolveError::BindingNotFound { .. } => {
682 CliError::new(ExitKind::NotFound, "PROJECTION_NOT_FOUND", message)
683 }
684 ResolveError::MalformedProjectionRef { .. } => {
685 CliError::new(ExitKind::Validation, "PROJECTION_INVALID_NAME", message)
686 }
687 // A scope nothing can interpret is a declaration defect the
688 // author must fix — Validation, with the legal forms named in
689 // the message rather than left for the reader to find.
690 ResolveError::UninterpretableScope { .. } => CliError::new(
691 ExitKind::Validation,
692 "PROJECTION_SCOPE_UNINTERPRETABLE",
693 message,
694 ),
695 },
696 };
697 mapped.with_details(json!({ "binding": binding_id }))
698}
699
700fn brief(ctx: &CliContext, args: BriefArgs) -> anyhow::Result<()> {
701 let (_shape, root) = ctx.workspace_shape().ok_or_else(|| {
702 workspace_not_initialised_error(
703 "not inside a Memstead workspace (no `.memstead/workspace.toml` in any ancestor)",
704 )
705 })?;
706
707 let cli_engine = ctx.cli_engine_at(&root)?;
708 let engine = cli_engine.base();
709
710 // Quarantine consult first: a binding whose stored file failed
711 // the load refuses typed with its reason (naming `projection
712 // migrate` for the legacy generations) instead of reporting
713 // not-found (agent-trust plan 04).
714 if let Some(binding_id) = args.binding.as_deref()
715 && let Ok(configs) = load_pipeline_configs(&root)
716 && configs
717 .quarantined
718 .iter()
719 .any(|q| format!("{}/{}", q.mem, q.name) == binding_id)
720 {
721 return Err(binding_miss_error(&configs, binding_id).into());
722 }
723
724 // Group-C briefs: verify / sync render for one named binding (no rotation).
725 // Both are read-only on the destination mem — the sync brief's repairs reach
726 // the mem only when an agent acts on it through the MCP mutation surface.
727 if args.verify || args.sync {
728 let binding_id = args.binding.ok_or_else(|| {
729 CliError::new(
730 ExitKind::Validation,
731 "PROJECTION_BRIEF_BINDING_REQUIRED",
732 format!(
733 "`projection brief --{}` needs a binding id `<mem>/<stem>` — it renders one \
734 binding's brief, not an `--all` rotation",
735 if args.verify { "verify" } else { "sync" }
736 ),
737 )
738 })?;
739 // D6/AC4 (backlog-sweep plan 03, decision 13): the sync brief is the
740 // maintenance-writer prompt — rendering it for a binding whose sync
741 // operation is not enabled spends a loop's work slot on work the
742 // engine will refuse to apply. Gate render exactly like `projection
743 // advance` does, with the one-command remedy in `details`.
744 if args.sync {
745 let configs = load_pipeline_configs(&root).map_err(|e| {
746 CliError::new(
747 ExitKind::Generic,
748 "PROJECTION_LOAD_FAILED",
749 format!("could not load binding store: {e}"),
750 )
751 .with_details(json!({ "error": e.to_string() }))
752 })?;
753 if let Some(record) = configs
754 .bindings
755 .iter()
756 .find(|r| format!("{}/{}", r.mem, r.name) == binding_id)
757 && record.config.operations.sync.is_none()
758 {
759 return Err(absent_sync_error(&binding_id, &record.config).into());
760 }
761 }
762 // Same consult verify makes, for the same reason: a quarantined
763 // destination serves no entities, so the rendered brief would describe
764 // a mem that is not there. `render_*_brief_for` also reaches
765 // `reconcile_exclusions`, which prunes authored exclusions whose
766 // sources appear to hold nothing — the 181 lost dispositions in the
767 // filed incident. Refuse above it, not after.
768 // Deliberately NOT an `if let … && let …` chain: an `Err` anywhere in
769 // the resolution would fall through to the unguarded render, which is
770 // the very path that reaches `reconcile_exclusions`. A guard that can
771 // silently skip is not a guard. Each step propagates its own typed
772 // failure instead, exactly as the render below would have.
773 let configs = load_pipeline_configs(&root).map_err(|e| {
774 CliError::new(
775 ExitKind::Generic,
776 "PROJECTION_LOAD_FAILED",
777 format!("could not load binding store: {e}"),
778 )
779 .with_details(json!({ "error": e.to_string() }))
780 })?;
781 let quarantine = match configs
782 .bindings
783 .iter()
784 .find(|r| format!("{}/{}", r.mem, r.name) == binding_id)
785 {
786 Some(record) => {
787 let resolved = resolve_binding_run(&binding_id, &record.config)
788 .map_err(|e| map_resolve_err(&binding_id, e))?;
789 engine
790 .quarantine_reason(&resolved.destination_mem)
791 .map(|q| (resolved.destination_mem.clone(), q.clone()))
792 }
793 // An unknown binding is the render's own refusal to make, with
794 // its nearest-name help; not this guard's.
795 None => None,
796 };
797 if let Some((destination_mem, q)) = quarantine {
798 return Err(CliError::new(
799 ExitKind::Validation,
800 "MEM_QUARANTINED",
801 format!(
802 "brief refused for `{binding_id}`: the destination mem `{}` is \
803quarantined ({}) — it serves no entities, so the brief would describe a mem \
804that is not there. Repair the mem, then re-run",
805 destination_mem, q.reason_message
806 ),
807 )
808 .with_details(json!({
809 "binding": binding_id,
810 "mem": destination_mem,
811 "reason_code": q.reason_code,
812 "reason": q.reason_message,
813 }))
814 .into());
815 }
816 let (rendered, operation) = if args.verify {
817 (
818 render_verify_brief_for(engine, &root, &binding_id),
819 OperationKind::Verify,
820 )
821 } else {
822 (
823 render_sync_brief_for(engine, &root, &binding_id),
824 OperationKind::Sync,
825 )
826 };
827 let rendered = rendered.map_err(|e| map_brief_err(&binding_id, e))?;
828
829 if ctx.json {
830 print_json(&json!({ "brief": rendered, "operation": operation.as_wire() }))?;
831 } else {
832 print!("{rendered}");
833 }
834 return Ok(());
835 }
836
837 // Resolve which (binding, operation) pair to render: a named binding
838 // (canonical `<mem>/<stem>`, build), or the next due pair in a round-robin
839 // `--all` rotation (which advances the cursor + backoff state). The
840 // rotation's operation set is `--operation` (default: build only — the
841 // classic rotation, byte-stable for existing callers).
842 let (selected, never_rotated) = match args.binding {
843 Some(binding) if !args.all => (Some((binding, OperationKind::Build)), Vec::new()),
844 _ => {
845 let configs = load_pipeline_configs(&root).map_err(|e| {
846 CliError::new(
847 ExitKind::Generic,
848 "PROJECTION_LOAD_FAILED",
849 format!("could not load binding store: {e}"),
850 )
851 .with_details(json!({ "error": e.to_string() }))
852 })?;
853 // Distinguish "nothing is configured" from "everything is backing
854 // off". Both otherwise collapse into the same `None` from
855 // `select_next_due_operation`, but the two outcomes want different
856 // caller responses: an empty store is a setup prompt, a
857 // backing-off pass is a no-op retry. Emit the empty-store signal
858 // explicitly so a caller (the plugin router, a status display) can
859 // branch on it.
860 if configs.bindings.is_empty() {
861 if ctx.json {
862 print_json(&json!({ "no_bindings": true }))?;
863 } else {
864 println!("> **[projection] No bindings configured in this workspace yet.**");
865 }
866 return Ok(());
867 }
868 // Criterion 4's --all half (backlog-sweep plan 03): the rotation
869 // drops pairs the binding never loop-declared — by design
870 // (consent lives in the declaration), but never silently. Say
871 // which (binding, op) pairs the requested filter can never
872 // rotate: JSON carries them structurally; markdown puts the note
873 // on stderr so the brief on stdout stays a verbatim prompt.
874 let never_rotated = not_loop_declared(&configs, args.operation.to_filter());
875 if !ctx.json {
876 for (binding, op) in &never_rotated {
877 eprintln!(
878 "[projection] skipped from rotation: `{binding}` declares no \
879 loop-triggered {} operation (enable with `memstead projection \
880 enable {} {binding}`)",
881 op.as_wire(),
882 op.as_wire(),
883 );
884 }
885 }
886 let picked = select_next_due_operation(
887 engine,
888 &root,
889 &configs,
890 args.operation.to_filter(),
891 args.consume,
892 );
893 (picked, never_rotated)
894 }
895 };
896 let not_rotated_json = never_rotated
897 .iter()
898 .map(|(b, o)| json!({ "binding": b, "operation": o.as_wire() }))
899 .collect::<Vec<_>>();
900
901 let Some((binding_id, operation)) = selected else {
902 // Every eligible pair is backing off (or not due) this pass — a valid
903 // outcome, the loop's quiet yield.
904 if ctx.json {
905 print_json(&json!({ "skipped": true, "not_rotated": not_rotated_json }))?;
906 } else {
907 println!(
908 "> **[projection] Skipped — every eligible binding is backing off this pass.**"
909 );
910 }
911 return Ok(());
912 };
913
914 // Dispatch to the selected operation's renderer: the rotation hands back
915 // build / sync / verify pairs, each with its own brief.
916 let rendered = match operation {
917 OperationKind::Build => render_ingest_brief(engine, &root, &binding_id, args.consume),
918 OperationKind::Sync => render_sync_brief_for(engine, &root, &binding_id),
919 OperationKind::Verify => render_verify_brief_for(engine, &root, &binding_id),
920 }
921 .map_err(|e| map_brief_err(&binding_id, e))?;
922
923 if ctx.json {
924 print_json(&json!({
925 "brief": rendered,
926 "operation": operation.as_wire(),
927 "not_rotated": not_rotated_json,
928 }))?;
929 } else {
930 // The brief *is* the stdout content (the skill pipes it as the agent
931 // prompt) — write it verbatim, no added trailing newline.
932 print!("{rendered}");
933 }
934 Ok(())
935}
936
937/// Is `value` a single, plain path component — safe to use verbatim as a `<mem>`
938/// or `<stem>` dir/file segment and as half of the binding id? Mirrors
939/// `pipeline_store`'s internal component guard so `init` refuses with a clear
940/// typed code up front rather than surfacing a store IO error mid-scaffold.
941fn is_single_component(value: &str) -> bool {
942 !value.is_empty()
943 && value != "."
944 && value != ".."
945 && !value.contains('/')
946 && !value.contains('\\')
947 && !value.contains(':')
948 && !value.contains('\0')
949}
950
951/// Derive a binding stem from a `--source` pointer: its final path component
952/// (trailing slashes trimmed). `../public` → `public`; `home` → `home`;
953/// `https://example.com/manual` → `manual`.
954fn derive_stem(source: &str) -> String {
955 source
956 .trim_end_matches('/')
957 .rsplit('/')
958 .next()
959 .unwrap_or(source)
960 .to_string()
961}
962
963/// Map a store write failure during scaffolding to a typed CLI error.
964fn init_write_error(binding_id: &str, err: StoreError) -> CliError {
965 CliError::new(
966 ExitKind::Generic,
967 "PROJECTION_INIT_FAILED",
968 format!("could not scaffold binding `{binding_id}`: {err}"),
969 )
970 .with_details(json!({ "binding": binding_id, "error": err.to_string() }))
971}
972
973fn init(ctx: &CliContext, args: InitArgs) -> anyhow::Result<()> {
974 let (_shape, root) = ctx.workspace_shape().ok_or_else(|| {
975 workspace_not_initialised_error(
976 "not inside a Memstead workspace (no `.memstead/workspace.toml` in any ancestor)",
977 )
978 })?;
979
980 let mem = args.mem;
981 let stem = args
982 .name
983 .clone()
984 .unwrap_or_else(|| derive_stem(&args.source));
985
986 // `mem` and `stem` become three file-path components and the binding id —
987 // refuse anything that is not a single plain component before touching disk.
988 for (kind, value) in [("mem", mem.as_str()), ("name", stem.as_str())] {
989 if !is_single_component(value) {
990 return Err(CliError::new(
991 ExitKind::Validation,
992 "PROJECTION_INVALID_NAME",
993 format!(
994 "invalid {kind} '{}': must be a single path component (no separators, \
995 traversal segments, ':' or NUL) — pass an explicit --name",
996 value.escape_default()
997 ),
998 )
999 .with_details(json!({ "kind": kind, "value": value }))
1000 .into());
1001 }
1002 }
1003
1004 let binding_id = format!("{mem}/{stem}");
1005 let medium_type = args.medium_type.to_medium_type();
1006
1007 // Refuse — without touching disk — when a binding of this id already exists
1008 // (D8: `init` never overwrites). The binding occupies the per-mem
1009 // projections tier; its presence is the id-collision signal.
1010 let binding_path = root
1011 .join(".memstead")
1012 .join("projections")
1013 .join(&mem)
1014 .join(format!("{stem}.json"));
1015 if binding_path.exists() {
1016 return Err(CliError::new(
1017 ExitKind::Validation,
1018 "PROJECTION_EXISTS",
1019 format!(
1020 "a binding `{binding_id}` already exists at \
1021 .memstead/projections/{mem}/{stem}.json — `projection init` never overwrites; \
1022 choose a different --name or edit the existing binding"
1023 ),
1024 )
1025 .with_details(json!({ "binding": binding_id }))
1026 .into());
1027 }
1028
1029 // The scaffolded record comes from the engine — one definition of "a
1030 // fresh binding" (source scoped `**/*`, materialised deny defaults,
1031 // matrix-filtered operations, prune where sync survived), shared with
1032 // every other front door that creates one.
1033 let scaffolded = memstead_base::binding::scaffold_binding(ScaffoldParams {
1034 destination_mem: &mem,
1035 source_name: &stem,
1036 pointer: &args.source,
1037 medium_type,
1038 intent: args.intent.clone(),
1039 additional_deny_paths: Vec::new(),
1040 });
1041 let binding = scaffolded.binding;
1042 let operations = scaffolded.operations;
1043 let mut warnings = scaffolded.warnings;
1044
1045 // Out-of-workspace medium base — a supported shape with one honest
1046 // caveat, named NOW, at the layout decision. The operation still succeeds.
1047 if let Some(w) =
1048 memstead_base::ingest::cursor::out_of_root_layout_warning(&args.source, &root, medium_type)
1049 {
1050 warnings.push(w);
1051 }
1052
1053 // A source that is not there yet is legal to declare — the tree may
1054 // arrive later — but silence here is how the setup ramp produces a
1055 // binding that can never yield anything from a mistyped answer. Say it
1056 // at the moment the pointer is chosen, where the typo is still in view.
1057 if matches!(
1058 medium_type,
1059 memstead_base::MediumType::Codebase
1060 | memstead_base::MediumType::Filesystem
1061 | memstead_base::MediumType::Git
1062 ) {
1063 let base = memstead_base::ingest::cursor::medium_base(&args.source, &root);
1064 if !base.exists() {
1065 warnings.push(format!(
1066 "source '{}' resolves to '{}', which does not exist — the binding is \
1067 declared, but nothing can be read from it until that path is there \
1068 (check the path, or correct the pointer in \
1069 .memstead/projections/{mem}/{stem}.json)",
1070 args.source,
1071 base.display(),
1072 ));
1073 }
1074 }
1075
1076 // Write the one record. The id-collision refusal above already
1077 // guaranteed a fresh binding, so this path only runs on a clean
1078 // scaffold; a store IO failure surfaces the typed
1079 // `PROJECTION_INIT_FAILED`.
1080 write_binding(&root, &mem, &stem, &binding).map_err(|e| init_write_error(&binding_id, e))?;
1081
1082 let created = vec![format!(".memstead/projections/{mem}/{stem}.json")];
1083
1084 if ctx.json {
1085 // D8's pinned skill contract: { binding, created, operations, warnings }.
1086 print_json(&json!({
1087 "binding": binding_id,
1088 "created": created,
1089 "operations": operations,
1090 "warnings": warnings,
1091 }))?;
1092 } else {
1093 let mut out = format!("# Projection init\n\nScaffolded binding `{binding_id}`:\n");
1094 for c in &created {
1095 out.push_str(&format!("- `{c}`\n"));
1096 }
1097 out.push_str(&format!("\nOperations: {}\n", operations.join(", ")));
1098 if !warnings.is_empty() {
1099 out.push_str("\n## Warnings\n\n");
1100 for w in &warnings {
1101 out.push_str(&format!("- {w}\n"));
1102 }
1103 }
1104 print_markdown(&out);
1105 }
1106 Ok(())
1107}
1108
1109fn map_migrate_err(err: BindingMigrateError) -> CliError {
1110 // Spell each `PROJECTION_*` token as a literal at its own construction site
1111 // so the generated error index (xtask) picks them up — a variable `code`
1112 // is invisible to the string-literal scanner.
1113 let message = err.to_string();
1114 match &err {
1115 BindingMigrateError::RefinementModeDeleted { .. } => CliError::new(
1116 ExitKind::Validation,
1117 "PROJECTION_MIGRATE_REFINEMENT",
1118 message,
1119 ),
1120 BindingMigrateError::MalformedProjectionRef { .. } => CliError::new(
1121 ExitKind::Validation,
1122 "PROJECTION_MIGRATE_MALFORMED_REF",
1123 message,
1124 ),
1125 BindingMigrateError::DanglingProjectionRef { .. }
1126 | BindingMigrateError::DanglingFacetRef { .. }
1127 | BindingMigrateError::DanglingMediumRef { .. } => CliError::new(
1128 ExitKind::Validation,
1129 "PROJECTION_MIGRATE_DANGLING_REF",
1130 message,
1131 ),
1132 BindingMigrateError::OrphanRecords { .. } => CliError::new(
1133 ExitKind::Validation,
1134 "PROJECTION_MIGRATE_ORPHAN_RECORDS",
1135 message,
1136 ),
1137 }
1138}
1139
1140/// Does the workspace root carry a gen-1 legacy pipeline layout — the
1141/// pre-four-primitive `scopes|projections|ingests/` JSON folders at the root
1142/// (not under `.memstead/`)? Presence of any of the three marks it. This is the
1143/// trigger for folding the retired `pipeline migrate` conversion into
1144/// `projection migrate` (D10, gen-1 path).
1145fn has_legacy_root_layout(root: &std::path::Path) -> bool {
1146 ["scopes", "projections", "ingests"]
1147 .iter()
1148 .any(|d| root.join(d).is_dir())
1149}
1150
1151/// Map a store load failure during migrate to the typed generic code.
1152fn migrate_load_err(err: StoreError) -> CliError {
1153 CliError::new(
1154 ExitKind::Generic,
1155 "PROJECTION_MIGRATE_FAILED",
1156 format!("could not load pipeline config: {err}"),
1157 )
1158 .with_details(json!({ "error": err.to_string() }))
1159}
1160
1161/// Does the binding's `medium_pointer` (resolved against the workspace root)
1162/// point at the same location as a `reconcile-cursors.json` absolute key? Uses
1163/// canonicalization where both paths exist, else a lexical comparison (D10 —
1164/// "the binding whose medium pointer resolves to that path").
1165fn pointer_resolves_to(root: &std::path::Path, medium_pointer: &str, abs_path: &str) -> bool {
1166 let resolved = if medium_pointer.is_empty() {
1167 root.to_path_buf()
1168 } else {
1169 root.join(medium_pointer)
1170 };
1171 match (
1172 std::fs::canonicalize(&resolved),
1173 std::fs::canonicalize(abs_path),
1174 ) {
1175 (Ok(a), Ok(b)) => a == b,
1176 _ => resolved == std::path::Path::new(abs_path),
1177 }
1178}
1179
1180/// Scan `workspace.toml` for retired pipeline/cursor vocabulary. `projection
1181/// migrate` **never** writes `workspace.toml` (D10) — if it finds a stale
1182/// reference it returns a proposal block for the operator (or the migrating
1183/// session) to apply and commit explicitly, rather than rewriting it.
1184fn propose_workspace_toml(root: &std::path::Path) -> Option<String> {
1185 let path = root.join(".memstead").join("workspace.toml");
1186 let content = std::fs::read_to_string(path).ok()?;
1187 let hits: Vec<(usize, &str)> = content
1188 .lines()
1189 .enumerate()
1190 .filter(|(_, l)| {
1191 let low = l.to_lowercase();
1192 low.contains("reconcile-cursors") || low.contains("ingests/") || low.contains("ingest ")
1193 })
1194 .collect();
1195 if hits.is_empty() {
1196 return None;
1197 }
1198 let mut block = String::from(
1199 "## Proposal: workspace.toml (NOT applied)\n\n`projection migrate` never edits \
1200 `workspace.toml`. It found references to retired pipeline vocabulary — review and \
1201 update these lines by hand, then commit:\n\n",
1202 );
1203 for (i, line) in hits {
1204 block.push_str(&format!("- L{}: `{}`\n", i + 1, line.trim()));
1205 }
1206 Some(block)
1207}
1208
1209/// Binding-miss refusal that honours the quarantine roster
1210/// (agent-trust plan 04): a binding whose stored file failed the v2
1211/// version gate is QUARANTINED, not unknown — the refusal carries the
1212/// typed reason, whose message names `memstead projection migrate`
1213/// for the legacy generations. Healthy-miss keeps the historical
1214/// `PROJECTION_NOT_FOUND`.
1215fn binding_miss_error(configs: &memstead_base::BindingConfigs, binding_id: &str) -> CliError {
1216 if let Some(q) = configs
1217 .quarantined
1218 .iter()
1219 .find(|q| format!("{}/{}", q.mem, q.name) == binding_id)
1220 {
1221 return CliError::new(
1222 ExitKind::Validation,
1223 "PROJECTION_QUARANTINED",
1224 format!(
1225 "binding `{binding_id}` is quarantined — its stored file failed the load and \
1226 it serves no operations until repaired: [{}] {}",
1227 q.reason_code, q.reason_message
1228 ),
1229 )
1230 .with_details(json!({
1231 "binding": binding_id,
1232 "reason_code": q.reason_code,
1233 "reason_message": q.reason_message,
1234 "path": q.path,
1235 }));
1236 }
1237 CliError::new(
1238 ExitKind::NotFound,
1239 "PROJECTION_NOT_FOUND",
1240 format!(
1241 "no binding `{binding_id}` in this workspace — scaffold one with \
1242 `projection init` or migrate a legacy workspace with `projection migrate`"
1243 ),
1244 )
1245 .with_details(json!({ "binding": binding_id }))
1246}
1247
1248/// Consume a skill-written `reconcile-cursors.json` (D10/AC12): each
1249/// machine-absolute `"<mem>:<abs-path>": <sha>` entry seeds the `#synced`
1250/// baseline of every binding whose medium pointer resolves to that path (via
1251/// the engine's `set_mem_sync_state` writer — the engine owns mem-repo state),
1252/// then the file is **deleted** regardless of whether anything matched
1253/// (cursorless / unmatched bindings stay never-synced). Returns the seeded keys.
1254fn consume_reconcile_cursors(
1255 ctx: &CliContext,
1256 root: &std::path::Path,
1257) -> anyhow::Result<(Vec<String>, Option<String>)> {
1258 let cursor_path = root.join(".memstead").join("reconcile-cursors.json");
1259 if !cursor_path.exists() {
1260 return Ok((Vec::new(), None));
1261 }
1262 let cursors: std::collections::BTreeMap<String, String> = std::fs::read(&cursor_path)
1263 .ok()
1264 .and_then(|b| serde_json::from_slice(&b).ok())
1265 .unwrap_or_default();
1266
1267 let mut seeded: Vec<String> = Vec::new();
1268 if !cursors.is_empty() {
1269 let configs = load_pipeline_configs(root).map_err(migrate_load_err)?;
1270 // Repair-below-boot rule: cursor seeding needs a booted engine
1271 // (`set_mem_sync_state`), but `projection migrate` is itself a
1272 // named boot repair — when the workspace still does not boot
1273 // (e.g. a schema-pin failure alongside the projection
1274 // migration), seeding is explicitly DEFERRED rather than
1275 // deadlocking the repair verb or silently dropping the
1276 // cursors: the file is kept untouched and the notice names the
1277 // follow-up.
1278 let mut cli_engine = match ctx.cli_engine_at(root) {
1279 Ok(e) => e,
1280 Err(boot_err) => {
1281 return Ok((
1282 Vec::new(),
1283 Some(format!(
1284 "RECONCILE_CURSORS_DEFERRED: the workspace does not boot yet \
1285 ({boot_err:#}); reconcile-cursors.json was kept — repair the boot, \
1286 then re-run `memstead projection migrate` to seed the sync baselines"
1287 )),
1288 ));
1289 }
1290 };
1291 let engine = cli_engine.base_mut();
1292 for (cursor_key, sha) in &cursors {
1293 // Key is `"<mem>:<abs-path>"` — split on the first ':'.
1294 let Some((_cursor_mem, abs_path)) = cursor_key.split_once(':') else {
1295 continue;
1296 };
1297 for record in &configs.bindings {
1298 let binding_id = format!("{}/{}", record.mem, record.name);
1299 let Ok(resolved) = resolve_binding_run(&binding_id, &record.config) else {
1300 continue;
1301 };
1302 for source in &resolved.sources {
1303 if let ResolvedSource::Primary(p) = source
1304 && pointer_resolves_to(root, &p.pointer, abs_path)
1305 {
1306 let key = format!("{binding_id}/{}#synced", p.name);
1307 // `.is_ok()` dropped the outcome and with it any
1308 // report that a sibling had moved the config
1309 // (04/03, criterion 3). A seeding pass that silently
1310 // merged over someone is worth one line of output.
1311 if let Ok(outcome) = engine.set_mem_sync_state(
1312 &resolved.destination_mem,
1313 &key,
1314 sha,
1315 Some("projection migrate: seeded from reconcile-cursors.json"),
1316 ) {
1317 for w in &outcome.warnings {
1318 crate::output::print_markdown(&format!("> {w}"));
1319 }
1320 seeded.push(key);
1321 }
1322 }
1323 }
1324 }
1325 }
1326 }
1327 // Consumed — delete regardless of matches (D10: the file is retired here).
1328 let _ = std::fs::remove_file(&cursor_path);
1329 Ok((seeded, None))
1330}
1331
1332fn migrate(ctx: &CliContext, args: MigrateArgs) -> anyhow::Result<()> {
1333 let (_shape, root) = ctx.workspace_shape().ok_or_else(|| {
1334 workspace_not_initialised_error(
1335 "not inside a Memstead workspace (no `.memstead/workspace.toml` in any ancestor)",
1336 )
1337 })?;
1338
1339 // Gen-1 root-folder layout (`scopes|projections|ingests/` at the workspace
1340 // root) — the pre-four-primitive generation the retired `pipeline migrate`
1341 // command handled. Fold it in: materialize it into the four-primitive
1342 // `.memstead/` store first (mediums + facets + projections + ingests),
1343 // then fold to v2 below in the same pass. `--dry-run` reads the
1344 // root-folder configs directly without writing anything.
1345 let gen1 = has_legacy_root_layout(&root);
1346 if gen1 && !args.dry_run {
1347 migrate_legacy_pipeline(&root).map_err(|e| {
1348 CliError::new(
1349 ExitKind::Generic,
1350 "PROJECTION_MIGRATE_FAILED",
1351 format!("could not convert root-folder (gen-1) pipeline layout: {e}"),
1352 )
1353 .with_details(json!({ "error": e.to_string() }))
1354 })?;
1355 }
1356
1357 let configs = if gen1 && args.dry_run {
1358 read_legacy_pipeline_configs(&root).map_err(migrate_load_err)?
1359 } else {
1360 load_legacy_pipeline_configs(&root).map_err(migrate_load_err)?
1361 };
1362
1363 // Pure transforms first: any refusal (refinement / dangling / malformed /
1364 // orphan) aborts before a single file is touched — the migration is
1365 // all-or-nothing.
1366 //
1367 // Leg A (gen-2): merge each flat ingest into its projection and fold the
1368 // referenced facets + mediums inline — one v2 record per pipeline.
1369 let mut migrated = migrate_gen2_bindings(&configs).map_err(map_migrate_err)?;
1370
1371 // Leg B (v1 → v2): fold every on-disk `version: 1` binding of the
1372 // retired three-file store the same way, in place. Source names are the
1373 // facet names byte-verbatim, so sync watermarks keep resolving. A
1374 // version-less projection file no ingest schedules is inert leftovers —
1375 // refused with a remedy rather than silently dropped or left to break
1376 // the loader. (Skipped in the gen-1 dry-run, which previews in-memory.)
1377 let mut already_v2 = 0usize;
1378 if !(gen1 && args.dry_run) {
1379 let generations = load_projection_generations(&root).map_err(migrate_load_err)?;
1380 for (mem, name, generation) in generations {
1381 let binding_id = format!("{mem}/{name}");
1382 match generation {
1383 ProjectionGeneration::V2 => already_v2 += 1,
1384 ProjectionGeneration::V1(v1) => {
1385 let consumed = v1.source_facets.clone();
1386 let binding = fold_v1_binding(&binding_id, &mem, v1.as_ref(), &configs)
1387 .map_err(map_migrate_err)?;
1388 migrated.push(memstead_base::binding_migrate::MigratedBinding {
1389 id: binding_id,
1390 mem,
1391 name,
1392 ingest_name: String::new(),
1393 consumed_facets: consumed,
1394 binding,
1395 notes: Vec::new(),
1396 });
1397 }
1398 ProjectionGeneration::VersionLess => {
1399 if !migrated.iter().any(|m| m.mem == mem && m.name == name) {
1400 return Err(CliError::new(
1401 ExitKind::Validation,
1402 "PROJECTION_MIGRATE_INERT_PROJECTION",
1403 format!(
1404 "projection `{binding_id}` is a version-less gen-2 file no \
1405 ingest schedules — inert leftovers the loader refuses; delete \
1406 .memstead/projections/{mem}/{name}.json (or add an ingest) and \
1407 re-run `projection migrate`"
1408 ),
1409 )
1410 .with_details(json!({ "binding": binding_id }))
1411 .into());
1412 }
1413 }
1414 }
1415 }
1416 migrated.sort_by(|a, b| a.id.cmp(&b.id));
1417
1418 // Every medium/facet record must have folded into some binding —
1419 // an orphan would be silently dropped by the tree removal, so the
1420 // whole migration refuses instead, naming each leftover.
1421 let consumed: Vec<(String, String)> = migrated
1422 .iter()
1423 .flat_map(|m| m.consumed_facets.iter().map(|f| (m.mem.clone(), f.clone())))
1424 .collect();
1425 check_all_consumed(&configs, &consumed).map_err(map_migrate_err)?;
1426 }
1427
1428 // Validate each produced binding against the capability matrix. A
1429 // capability refusal reflects a pre-existing config problem the binding
1430 // faithfully carries; surface it as a per-binding warning rather than
1431 // aborting the promotion. The folded v2 record validates directly — no
1432 // external resolution.
1433 let mut warnings: Vec<serde_json::Value> = Vec::new();
1434 for m in &migrated {
1435 if let Err(refusals) = validate_binding(&m.binding) {
1436 for r in refusals {
1437 warnings.push(json!({
1438 "binding": m.id,
1439 "kind": "capability",
1440 "message": r.to_string(),
1441 }));
1442 }
1443 }
1444 for note in &m.notes {
1445 warnings.push(json!({
1446 "binding": m.id,
1447 "kind": "note",
1448 "message": note,
1449 }));
1450 }
1451 }
1452
1453 // Emit to disk unless previewing: promote each projection file to its v2
1454 // binding in place, remove each consumed flat ingest, then remove the
1455 // emptied `mediums/` and `facets/` trees (every record folded — the
1456 // orphan check above guaranteed it).
1457 if !args.dry_run {
1458 for m in &migrated {
1459 write_binding(&root, &m.mem, &m.name, &m.binding).map_err(|e| {
1460 CliError::new(
1461 ExitKind::Generic,
1462 "PROJECTION_MIGRATE_FAILED",
1463 format!("could not write binding `{}`: {e}", m.id),
1464 )
1465 .with_details(json!({ "binding": m.id, "error": e.to_string() }))
1466 })?;
1467 if !m.ingest_name.is_empty() {
1468 delete_ingest(&root, &m.ingest_name).map_err(|e| {
1469 CliError::new(
1470 ExitKind::Generic,
1471 "PROJECTION_MIGRATE_FAILED",
1472 format!("could not remove merged ingest `{}`: {e}", m.ingest_name),
1473 )
1474 .with_details(json!({ "ingest": m.ingest_name, "error": e.to_string() }))
1475 })?;
1476 }
1477 }
1478 remove_mediums_and_facets_trees(&root).map_err(|e| {
1479 CliError::new(
1480 ExitKind::Generic,
1481 "PROJECTION_MIGRATE_FAILED",
1482 format!("could not remove the emptied mediums/facets trees: {e}"),
1483 )
1484 .with_details(json!({ "error": e.to_string() }))
1485 })?;
1486 }
1487
1488 // AC12/D10: consume `reconcile-cursors.json` (seed `#synced` baselines, then
1489 // delete it) and surface a `workspace.toml` proposal for any retired-vocab
1490 // references — never rewriting workspace.toml. Both are no-ops in `--dry-run`.
1491 let ((seeded, cursors_deferred), proposal) = if args.dry_run {
1492 ((Vec::new(), None), None)
1493 } else {
1494 (
1495 consume_reconcile_cursors(ctx, &root)?,
1496 propose_workspace_toml(&root),
1497 )
1498 };
1499
1500 let bindings: Vec<&str> = migrated.iter().map(|m| m.id.as_str()).collect();
1501 if ctx.json {
1502 print_json(&json!({
1503 "ok": true,
1504 "dry_run": args.dry_run,
1505 "migrated": migrated.len(),
1506 "already_v2": already_v2,
1507 "bindings": bindings,
1508 "warnings": warnings,
1509 "cursors_seeded": seeded,
1510 "cursors_deferred": cursors_deferred,
1511 "workspace_toml_proposal": proposal,
1512 }))?;
1513 } else {
1514 let verb = if args.dry_run {
1515 "Would migrate"
1516 } else {
1517 "Migrated"
1518 };
1519 let mut out = format!(
1520 "# Projection migration\n\n{verb} {} binding(s) to v2 ({already_v2} already v2):\n",
1521 migrated.len()
1522 );
1523 for id in &bindings {
1524 out.push_str(&format!("- `{id}`\n"));
1525 }
1526 if !warnings.is_empty() {
1527 out.push_str("\n## Warnings\n\n");
1528 for w in &warnings {
1529 out.push_str(&format!(
1530 "- [{}] `{}`: {}\n",
1531 w["kind"].as_str().unwrap_or(""),
1532 w["binding"].as_str().unwrap_or(""),
1533 w["message"].as_str().unwrap_or(""),
1534 ));
1535 }
1536 }
1537 if !seeded.is_empty() {
1538 out.push_str("\n## Baselines seeded from reconcile-cursors.json\n\n");
1539 for key in &seeded {
1540 out.push_str(&format!("- `{key}`\n"));
1541 }
1542 }
1543 if let Some(notice) = &cursors_deferred {
1544 out.push_str(&format!("\n## Reconcile cursors deferred\n\n{notice}\n"));
1545 }
1546 if let Some(block) = &proposal {
1547 out.push('\n');
1548 out.push_str(block);
1549 }
1550 if !args.dry_run {
1551 out.push_str(
1552 "\nEach projection file was converted to a v2 single-record binding in place \
1553 (medium + facet content folded inline, source names preserved verbatim); \
1554 merged ingests and the emptied mediums/ and facets/ trees were removed.\n",
1555 );
1556 }
1557 print_markdown(&out);
1558 }
1559 Ok(())
1560}
1561
1562/// A malformed binding id (not `<mem>/<stem>`, or a half that is not a single
1563/// plain path component) — the same shape guard `init` applies to its
1564/// scaffolded id, spelled here so the failure is typed before any disk touch.
1565fn invalid_binding_id(binding_id: &str) -> CliError {
1566 CliError::new(
1567 ExitKind::Validation,
1568 "PROJECTION_INVALID_NAME",
1569 format!(
1570 "invalid binding id '{}': expected `<mem>/<stem>` with each half a single path \
1571 component (no extra separators, traversal segments, ':' or NUL)",
1572 binding_id.escape_default()
1573 ),
1574 )
1575 .with_details(json!({ "binding": binding_id }))
1576}
1577
1578/// Map a store IO/parse failure while enabling to a typed CLI error. The
1579/// missing-binding case is handled separately (existence pre-check →
1580/// `PROJECTION_NOT_FOUND`); this covers a present-but-unreadable/unparseable
1581/// binding file and write failures.
1582fn enable_failed(binding_id: &str, err: StoreError) -> CliError {
1583 CliError::new(
1584 ExitKind::Generic,
1585 "PROJECTION_ENABLE_FAILED",
1586 format!("could not enable operation on binding `{binding_id}`: {err}"),
1587 )
1588 .with_details(json!({ "binding": binding_id, "error": err.to_string() }))
1589}
1590
1591fn enable(ctx: &CliContext, args: EnableArgs) -> anyhow::Result<()> {
1592 let (_shape, root) = ctx.workspace_shape().ok_or_else(|| {
1593 workspace_not_initialised_error(
1594 "not inside a Memstead workspace (no `.memstead/workspace.toml` in any ancestor)",
1595 )
1596 })?;
1597
1598 let binding_id = args.binding;
1599 let op = args.operation;
1600
1601 // Parse the binding id `<mem>/<stem>`; refuse a malformed shape (or a half
1602 // that is not a single plain path component) before touching disk. Own the
1603 // halves so `binding_id` is free to move into JSON payloads later.
1604 let (mem, stem) = binding_id
1605 .split_once('/')
1606 .filter(|(m, n)| !m.is_empty() && !n.is_empty())
1607 .filter(|(m, n)| is_single_component(m) && is_single_component(n))
1608 .ok_or_else(|| invalid_binding_id(&binding_id))?;
1609 let mem = mem.to_string();
1610 let stem = stem.to_string();
1611
1612 // Missing binding file → PROJECTION_NOT_FOUND (NotFound exit). A present-
1613 // but-unparseable file is kept apart (→ PROJECTION_ENABLE_FAILED) by this
1614 // existence pre-check.
1615 let binding_path = root
1616 .join(".memstead")
1617 .join("projections")
1618 .join(&mem)
1619 .join(format!("{stem}.json"));
1620 if !binding_path.exists() {
1621 return Err(CliError::new(
1622 ExitKind::NotFound,
1623 "PROJECTION_NOT_FOUND",
1624 format!(
1625 "no binding `{binding_id}` at .memstead/projections/{mem}/{stem}.json — \
1626 scaffold one with `projection init` or migrate a legacy workspace with \
1627 `projection migrate`"
1628 ),
1629 )
1630 .with_details(json!({ "binding": binding_id }))
1631 .into());
1632 }
1633 // Quarantine consult before the raw read: a legacy/corrupt file
1634 // refuses with its typed reason (naming `projection migrate` for
1635 // the legacy generations) rather than a generic enable failure.
1636 if let Ok(configs) = load_pipeline_configs(&root)
1637 && configs
1638 .quarantined
1639 .iter()
1640 .any(|q| format!("{}/{}", q.mem, q.name) == binding_id)
1641 {
1642 return Err(binding_miss_error(&configs, &binding_id).into());
1643 }
1644 let mut binding =
1645 read_binding(&root, &mem, &stem).map_err(|e| enable_failed(&binding_id, e))?;
1646
1647 // Already present? Refuse without a partial write. Every operation block is
1648 // optional now (D1/AC4), so `build` is enableable too (the remedy a
1649 // build-less binding's brief refusal cites).
1650 let already = match op {
1651 EnableOperationArg::Build => binding.operations.build.is_some(),
1652 EnableOperationArg::Sync => binding.operations.sync.is_some(),
1653 EnableOperationArg::Verify => binding.operations.verify.is_some(),
1654 };
1655 if already {
1656 return Err(CliError::new(
1657 ExitKind::Validation,
1658 "PROJECTION_OP_ALREADY_ENABLED",
1659 format!(
1660 "operation `{}` is already enabled on binding `{binding_id}` — nothing to do",
1661 op.name()
1662 ),
1663 )
1664 .with_details(json!({ "binding": binding_id, "operation": op.name() }))
1665 .into());
1666 }
1667
1668 // Add the operation block with sensible defaults: `batch_size` mirrors the
1669 // build op's when present, else 20. Sync/verify default `trigger: manual`;
1670 // build defaults to a discovery/loop schedule (the common obligation shape).
1671 let batch_size = binding
1672 .operations
1673 .build
1674 .as_ref()
1675 .map_or(20, |b| b.batch_size);
1676 match op {
1677 EnableOperationArg::Build => {
1678 binding.operations.build = Some(BuildOperation {
1679 mode: BuildMode::Discovery,
1680 trigger: IngestTrigger::Loop,
1681 batch_size,
1682 post_actions: None,
1683 });
1684 }
1685 EnableOperationArg::Sync => {
1686 binding.operations.sync = Some(SyncOperation {
1687 trigger: IngestTrigger::Manual,
1688 batch_size,
1689 });
1690 }
1691 EnableOperationArg::Verify => {
1692 binding.operations.verify = Some(VerifyOperation {
1693 trigger: IngestTrigger::Manual,
1694 batch_size,
1695 adjudication_cap: DEFAULT_ADJUDICATION_CAP,
1696 full_resync_every: DEFAULT_FULL_RESYNC_EVERY,
1697 });
1698 }
1699 }
1700
1701 // Matrix validation: the v2 record carries its sources inline, so the
1702 // candidate validates directly — refuse if a source's medium half cannot
1703 // support the operation being enabled (e.g. `sync`/`verify` over a `web`
1704 // source). Refusals about *other* operations reflect pre-existing config
1705 // and do not block this enable (mirrors `migrate`'s treat-as-warning
1706 // posture). No write on refusal — the file stays byte-identical.
1707 if let Err(refusals) = validate_binding(&binding)
1708 && let Some(err) = refusals.iter().find(|r| {
1709 matches!(
1710 r,
1711 CapabilityError::OperationOutOfScope { operation, .. } if *operation == op.name()
1712 )
1713 })
1714 {
1715 return Err(CliError::new(
1716 ExitKind::Validation,
1717 "PROJECTION_CAPABILITY_UNSUPPORTED",
1718 err.to_string(),
1719 )
1720 .with_details(json!({ "binding": binding_id, "operation": op.name() }))
1721 .into());
1722 }
1723
1724 write_binding(&root, &mem, &stem, &binding).map_err(|e| enable_failed(&binding_id, e))?;
1725
1726 let mut operations: Vec<&str> = Vec::new();
1727 if binding.operations.build.is_some() {
1728 operations.push("build");
1729 }
1730 if binding.operations.sync.is_some() {
1731 operations.push("sync");
1732 }
1733 if binding.operations.verify.is_some() {
1734 operations.push("verify");
1735 }
1736
1737 if ctx.json {
1738 print_json(&json!({
1739 "binding": binding_id,
1740 "enabled": op.name(),
1741 "operations": operations,
1742 }))?;
1743 } else {
1744 print_markdown(&format!(
1745 "# Projection enable\n\nEnabled `{}` on binding `{binding_id}`.\n\nOperations: {}\n",
1746 op.name(),
1747 operations.join(", ")
1748 ));
1749 }
1750 Ok(())
1751}
1752
1753/// `projection edit` — the general binding-field patch over the shared
1754/// `pipeline_edit` layer. The layer owns everything that matters: patch
1755/// semantics (absent preserved, `null` clears, blocks replace whole),
1756/// engine-managed `version`, and validate-before-write with the
1757/// introduced-refusals-only rule. This function is command surface: id
1758/// grammar, quarantine consult, error translation, rendering.
1759fn edit(ctx: &CliContext, args: EditArgs) -> anyhow::Result<()> {
1760 let (_shape, root) = ctx.workspace_shape().ok_or_else(|| {
1761 workspace_not_initialised_error(
1762 "not inside a Memstead workspace (no `.memstead/workspace.toml` in any ancestor)",
1763 )
1764 })?;
1765
1766 let binding_id = args.binding;
1767 let (mem, stem) = binding_id
1768 .split_once('/')
1769 .filter(|(m, n)| !m.is_empty() && !n.is_empty())
1770 .filter(|(m, n)| is_single_component(m) && is_single_component(n))
1771 .ok_or_else(|| invalid_binding_id(&binding_id))?;
1772 let mem = mem.to_string();
1773 let stem = stem.to_string();
1774
1775 // Quarantine consult before the edit: a legacy/corrupt record refuses
1776 // with its typed reason (naming `projection migrate`) rather than a
1777 // generic not-found or parse failure.
1778 if let Ok(configs) = load_pipeline_configs(&root)
1779 && configs
1780 .quarantined
1781 .iter()
1782 .any(|q| format!("{}/{}", q.mem, q.name) == binding_id)
1783 {
1784 return Err(binding_miss_error(&configs, &binding_id).into());
1785 }
1786
1787 let binding =
1788 memstead_base::pipeline_edit::update_binding_json(&root, &mem, &stem, &args.patch)
1789 .map_err(|e| edit_refused(&binding_id, e))?;
1790
1791 let source_names: Vec<&str> = binding.sources.iter().map(|s| s.name.as_str()).collect();
1792 let mut operations: Vec<&str> = Vec::new();
1793 if binding.operations.build.is_some() {
1794 operations.push("build");
1795 }
1796 if binding.operations.sync.is_some() {
1797 operations.push("sync");
1798 }
1799 if binding.operations.verify.is_some() {
1800 operations.push("verify");
1801 }
1802
1803 if ctx.json {
1804 print_json(&json!({
1805 "binding": binding_id,
1806 "edited": true,
1807 "sources": source_names,
1808 "operations": operations,
1809 "record": serde_json::to_value(&binding)?,
1810 }))?;
1811 } else {
1812 print_markdown(&format!(
1813 "# Projection edit\n\nPatched binding `{binding_id}`.\n\nSources: {}\nOperations: {}\n",
1814 source_names.join(", "),
1815 operations.join(", ")
1816 ));
1817 }
1818 Ok(())
1819}
1820
1821/// Translate a [`pipeline_edit`] refusal into the CLI's typed envelope.
1822/// Nothing was written on any of these — the layer validates before it
1823/// writes.
1824fn edit_refused(
1825 binding_id: &str,
1826 err: memstead_base::pipeline_edit::PipelineEditError,
1827) -> CliError {
1828 use memstead_base::pipeline_edit::PipelineEditError as E;
1829 match &err {
1830 E::NotFound { .. } => CliError::new(
1831 ExitKind::NotFound,
1832 "PROJECTION_NOT_FOUND",
1833 format!(
1834 "no binding `{binding_id}` — scaffold one with `projection init` or migrate a \
1835 legacy workspace with `projection migrate`"
1836 ),
1837 )
1838 .with_details(json!({ "binding": binding_id })),
1839 E::InvalidJson { message, .. } => CliError::new(
1840 ExitKind::Validation,
1841 "PROJECTION_EDIT_INVALID_JSON",
1842 format!("the patch for `{binding_id}` did not deserialize: {message}"),
1843 )
1844 .with_details(json!({ "binding": binding_id, "error": message })),
1845 E::Capability { refusals, .. } => CliError::new(
1846 ExitKind::Validation,
1847 "PROJECTION_EDIT_REFUSED",
1848 format!(
1849 "the patch would introduce validation refusals on `{binding_id}` (nothing was \
1850 written): {}",
1851 refusals
1852 .iter()
1853 .map(|r| r.to_string())
1854 .collect::<Vec<_>>()
1855 .join("; ")
1856 ),
1857 )
1858 .with_details(json!({
1859 "binding": binding_id,
1860 "refusals": refusals.iter().map(|r| r.to_string()).collect::<Vec<_>>(),
1861 })),
1862 _ => CliError::new(
1863 ExitKind::Generic,
1864 "PROJECTION_EDIT_FAILED",
1865 format!("could not edit binding `{binding_id}`: {err}"),
1866 )
1867 .with_details(json!({ "binding": binding_id, "error": err.to_string() })),
1868 }
1869}
1870
1871/// `projection check-path` — deny verdicts for tool-call candidates, engine-
1872/// free: binding records are pure file I/O and the dialect evaluation needs
1873/// no store, so the check stays cheap enough to run on every tool call.
1874fn check_path(ctx: &CliContext, args: CheckPathArgs) -> anyhow::Result<()> {
1875 let (_shape, root) = ctx.workspace_shape().ok_or_else(|| {
1876 workspace_not_initialised_error(
1877 "not inside a Memstead workspace (no `.memstead/workspace.toml` in any ancestor)",
1878 )
1879 })?;
1880
1881 // Batch payload first: its `cwd` outranks the flag, and a malformed
1882 // payload refuses before any binding work — never a part-answer.
1883 let (candidates, payload_cwd): (Vec<String>, Option<std::path::PathBuf>) = if args.batch {
1884 let mut raw = String::new();
1885 std::io::Read::read_to_string(&mut std::io::stdin(), &mut raw)
1886 .map_err(|e| batch_invalid(format!("stdin unreadable: {e}")))?;
1887 let value: serde_json::Value =
1888 serde_json::from_str(&raw).map_err(|e| batch_invalid(format!("not JSON: {e}")))?;
1889 let paths = value
1890 .get("paths")
1891 .and_then(|p| p.as_array())
1892 .ok_or_else(|| batch_invalid("missing `paths` array".to_string()))?;
1893 let mut out = Vec::with_capacity(paths.len());
1894 for p in paths {
1895 match p.as_str() {
1896 Some(s) => out.push(s.to_string()),
1897 None => {
1898 return Err(batch_invalid(format!("non-string entry in `paths`: {p}")).into());
1899 }
1900 }
1901 }
1902 let cwd = match value.get("cwd") {
1903 None | Some(serde_json::Value::Null) => None,
1904 Some(serde_json::Value::String(s)) => Some(std::path::PathBuf::from(s)),
1905 Some(other) => {
1906 return Err(batch_invalid(format!("`cwd` is not a string: {other}")).into());
1907 }
1908 };
1909 (out, cwd)
1910 } else {
1911 (vec![args.path.clone().expect("clap requires PATH")], None)
1912 };
1913
1914 // The binding: named, or the active one (published by the last consuming
1915 // brief render). No active binding is a typed refusal, never an implicit
1916 // "allowed" — the caller decides that an unanswerable check fails open.
1917 let binding_id = match args.binding.clone() {
1918 Some(id) => id,
1919 None => memstead_base::ingest::read_active_binding_file(&root).ok_or_else(|| {
1920 CliError::new(
1921 ExitKind::NotFound,
1922 "NO_ACTIVE_BINDING",
1923 "no active binding — no consuming brief render has published one; name a \
1924 binding explicitly with --binding <mem>/<stem>",
1925 )
1926 })?,
1927 };
1928
1929 let configs = load_pipeline_configs(&root).map_err(|e| {
1930 CliError::new(
1931 ExitKind::Generic,
1932 "PROJECTION_LOAD_FAILED",
1933 format!("binding store unreadable: {e}"),
1934 )
1935 })?;
1936 let binding = configs
1937 .bindings
1938 .iter()
1939 .find(|r| format!("{}/{}", r.mem, r.name) == binding_id)
1940 .map(|r| &r.config)
1941 .ok_or_else(|| binding_miss_error(&configs, &binding_id))?;
1942
1943 let cwd = match payload_cwd.or(args.cwd) {
1944 Some(dir) => dir,
1945 None => std::env::current_dir()?,
1946 };
1947 let verdicts =
1948 memstead_base::ingest::check_deny_paths(&binding.deny_paths, &candidates, &cwd, &root);
1949
1950 if ctx.json {
1951 print_json(&json!({
1952 "binding": binding_id,
1953 "results": verdicts
1954 .iter()
1955 .map(|v| json!({
1956 "path": v.path,
1957 "denied": v.denied,
1958 "matched": v.matched,
1959 }))
1960 .collect::<Vec<_>>(),
1961 }))?;
1962 } else {
1963 let mut out = format!("# Check path — binding `{binding_id}`\n\n");
1964 for v in &verdicts {
1965 match &v.matched {
1966 Some(entry) => {
1967 out.push_str(&format!("- DENIED `{}` — matched `{entry}`\n", v.path));
1968 }
1969 None => out.push_str(&format!("- allowed `{}`\n", v.path)),
1970 }
1971 }
1972 print_markdown(&out);
1973 }
1974 Ok(())
1975}
1976
1977/// A malformed `--batch` payload: refuse whole, name the defect.
1978fn batch_invalid(reason: String) -> CliError {
1979 CliError::new(
1980 ExitKind::Validation,
1981 "INVALID_INPUT",
1982 format!(
1983 "--batch expects one JSON object on stdin — {{\"cwd\": \"<dir>\", \
1984 \"paths\": [\"...\"]}} — {reason}"
1985 ),
1986 )
1987}
1988
1989/// Map a `resolve_binding_run` failure to a typed CLI error. With inline
1990/// sources the dangling facet/medium refusals are gone; a malformed id is the
1991/// Validation-shaped name error, everything else generic.
1992fn map_resolve_err(binding_id: &str, err: ResolveError) -> CliError {
1993 let message = err.to_string();
1994 let mapped = match err {
1995 ResolveError::MalformedProjectionRef { .. } => {
1996 CliError::new(ExitKind::Validation, "PROJECTION_INVALID_NAME", message)
1997 }
1998 ResolveError::UninterpretableScope { .. } => CliError::new(
1999 ExitKind::Validation,
2000 "PROJECTION_SCOPE_UNINTERPRETABLE",
2001 message,
2002 ),
2003 _ => CliError::new(ExitKind::Generic, "PROJECTION_ADVANCE_FAILED", message),
2004 };
2005 mapped.with_details(json!({ "binding": binding_id }))
2006}
2007
2008/// Map an [`AdvanceError`] to a typed CLI error. The unknown-artifact refusal
2009/// is the D7 gate (Validation); a malformed id is a Validation-shaped name
2010/// error; store / engine failures are generic. Codes are spelled as literals at
2011/// each site so the generated error index picks them up.
2012fn map_advance_err(binding_id: &str, err: AdvanceError) -> CliError {
2013 let message = err.to_string();
2014 match &err {
2015 AdvanceError::MalformedId(_) => {
2016 CliError::new(ExitKind::Validation, "PROJECTION_INVALID_NAME", message)
2017 .with_details(json!({ "binding": binding_id }))
2018 }
2019 AdvanceError::UnknownArtifact {
2020 artifacts,
2021 suggestions,
2022 ..
2023 } => {
2024 // `corrected_artifacts` maps each medium-relative-looking id to
2025 // the workspace-relative id the slice actually presented — the
2026 // machine-readable half of the message's remedy.
2027 let corrected: serde_json::Map<String, serde_json::Value> = suggestions
2028 .iter()
2029 .map(|(supplied, corrected)| {
2030 (
2031 supplied.clone(),
2032 serde_json::Value::String(corrected.clone()),
2033 )
2034 })
2035 .collect();
2036 CliError::new(
2037 ExitKind::Validation,
2038 "PROJECTION_ADVANCE_UNKNOWN_ARTIFACT",
2039 message,
2040 )
2041 .with_details(json!({
2042 "binding": binding_id,
2043 "unknown_artifacts": artifacts,
2044 "corrected_artifacts": corrected,
2045 }))
2046 }
2047 AdvanceError::Store(_) | AdvanceError::Engine(_) => {
2048 CliError::new(ExitKind::Generic, "PROJECTION_ADVANCE_FAILED", message)
2049 .with_details(json!({ "binding": binding_id }))
2050 }
2051 }
2052}
2053
2054fn advance(ctx: &CliContext, args: AdvanceArgs) -> anyhow::Result<()> {
2055 let (_shape, root) = ctx.workspace_shape().ok_or_else(|| {
2056 workspace_not_initialised_error(
2057 "not inside a Memstead workspace (no `.memstead/workspace.toml` in any ancestor)",
2058 )
2059 })?;
2060
2061 let binding_id = args.binding;
2062
2063 // Parse the dispositions payload up front — a malformed `--dispositions`
2064 // refuses cheaply (before loading configs or an engine) with a typed code.
2065 let dispositions: std::collections::BTreeMap<String, DispositionInput> =
2066 serde_json::from_str(&args.dispositions).map_err(|e| {
2067 CliError::new(
2068 ExitKind::Validation,
2069 "PROJECTION_INVALID_DISPOSITIONS",
2070 format!(
2071 "--dispositions must be a JSON object mapping artifact id → either a \
2072 disposition string (e.g. \"worked\") or an object \
2073 {{\"disposition\": \"excluded\", \"rationale\": \"...\"}}: {e}"
2074 ),
2075 )
2076 .with_details(json!({ "error": e.to_string() }))
2077 })?;
2078
2079 // Find the binding by canonical id in the v1 store.
2080 let configs = load_pipeline_configs(&root).map_err(|e| {
2081 CliError::new(
2082 ExitKind::Generic,
2083 "PROJECTION_ADVANCE_FAILED",
2084 format!("could not load pipeline config: {e}"),
2085 )
2086 .with_details(json!({ "error": e.to_string() }))
2087 })?;
2088 let record = configs
2089 .bindings
2090 .iter()
2091 .find(|r| format!("{}/{}", r.mem, r.name) == binding_id)
2092 .ok_or_else(|| binding_miss_error(&configs, &binding_id))?;
2093
2094 // D6/AC4: advance is the sync (maintenance-write) path — refuse when the
2095 // binding declares no `sync` operation, carrying the one-command remedy
2096 // `projection enable sync <binding>` (which, run verbatim, makes it
2097 // succeed) — except over a medium whose capability row refuses sync
2098 // outright, where the gap is named instead of a remedy that would bounce.
2099 if record.config.operations.sync.is_none() {
2100 return Err(absent_sync_error(&binding_id, &record.config).into());
2101 }
2102
2103 let resolved = resolve_binding_run(&binding_id, &record.config)
2104 .map_err(|e| map_resolve_err(&binding_id, e))?;
2105
2106 // The engine is mutable — a completing advance writes the `#synced`
2107 // baseline token through the sync-state writer.
2108 let mut cli_engine = ctx.cli_engine_at(&root)?;
2109 let engine = cli_engine.base_mut();
2110
2111 let outcome = advance_baseline(engine, &root, &resolved, &dispositions)
2112 .map_err(|e| map_advance_err(&binding_id, e))?;
2113
2114 if ctx.json {
2115 print_json(&json!({
2116 "binding": outcome.binding,
2117 "completed": outcome.completed,
2118 "disposed": outcome.disposed,
2119 "pending": outcome.pending,
2120 "remainder": outcome.remainder,
2121 "tokens_written": outcome.tokens_written,
2122 "warnings": outcome.warnings,
2123 }))?;
2124 } else {
2125 let mut out = format!(
2126 "# Projection advance\n\nBinding `{}`: {} artifact(s) disposed, {} remaining.\n",
2127 outcome.binding, outcome.disposed, outcome.pending
2128 );
2129 if outcome.completed {
2130 if outcome.disposed == 0 && outcome.pending == 0 {
2131 out.push_str(
2132 "\nNo artifacts were presented this pass — the sync baseline advanced.\n",
2133 );
2134 } else {
2135 out.push_str(
2136 "\nEvery presented artifact is disposed — the sync baseline advanced.\n",
2137 );
2138 }
2139 if !outcome.tokens_written.is_empty() {
2140 out.push_str("\nBaseline tokens written:\n");
2141 for key in &outcome.tokens_written {
2142 out.push_str(&format!("- `{key}`\n"));
2143 }
2144 }
2145 } else {
2146 out.push_str(
2147 "\nRemainder still pending — re-run `projection advance` after judging the rest \
2148 (a brief re-render shows what is left).\n",
2149 );
2150 }
2151 if !outcome.warnings.is_empty() {
2152 out.push_str("\n## Warnings\n\n");
2153 for w in &outcome.warnings {
2154 out.push_str(&format!("- {w}\n"));
2155 }
2156 }
2157 print_markdown(&out);
2158 }
2159 Ok(())
2160}
2161
2162/// Map an [`ExcludeError`] to a typed CLI error. The non-member refusal is the
2163/// S(D)-membership gate (Validation); a malformed id is a Validation-shaped name
2164/// error; store failures are generic. Codes are spelled as literals at each site
2165/// so the generated error index picks them up.
2166fn map_exclude_err(binding_id: &str, err: ExcludeError) -> CliError {
2167 let message = err.to_string();
2168 match &err {
2169 ExcludeError::MalformedId(_) => {
2170 CliError::new(ExitKind::Validation, "PROJECTION_INVALID_NAME", message)
2171 .with_details(json!({ "binding": binding_id }))
2172 }
2173 // The recovery IS the candidate list, so it rides `details` under a
2174 // key a caller can branch on, not only the prose.
2175 ExcludeError::AmbiguousArtifact { ambiguous } => CliError::new(
2176 ExitKind::Validation,
2177 "PROJECTION_EXCLUDE_AMBIGUOUS_ARTIFACT",
2178 message,
2179 )
2180 .with_details(json!({
2181 "binding": binding_id,
2182 "ambiguous": ambiguous,
2183 })),
2184 ExcludeError::NotSourceMember {
2185 artifacts, nearest, ..
2186 } => CliError::new(
2187 ExitKind::Validation,
2188 "PROJECTION_EXCLUDE_NOT_SOURCE_MEMBER",
2189 {
2190 let hints: Vec<String> = nearest
2191 .iter()
2192 .filter(|(_, ids)| !ids.is_empty())
2193 .map(|(id, ids)| format!("{id} → nearest known: {}", ids.join(", ")))
2194 .collect();
2195 if hints.is_empty() {
2196 message
2197 } else {
2198 format!("{message}; {}", hints.join("; "))
2199 }
2200 },
2201 )
2202 .with_details(json!({
2203 "binding": binding_id,
2204 "not_source_members": artifacts,
2205 // The nearest known ids per offender: the spelling to repair to.
2206 "nearest": nearest,
2207 })),
2208 ExcludeError::PartialEnumeration { facet, reason } => CliError::new(
2209 ExitKind::Validation,
2210 "PROJECTION_EXCLUDE_PARTIAL_ENUMERATION",
2211 message,
2212 )
2213 .with_details(json!({ "binding": binding_id, "facet": facet, "reason": reason })),
2214 ExcludeError::Store(_) => {
2215 CliError::new(ExitKind::Generic, "PROJECTION_EXCLUDE_FAILED", message)
2216 .with_details(json!({ "binding": binding_id }))
2217 }
2218 }
2219}
2220
2221fn exclude(ctx: &CliContext, args: ExcludeArgs) -> anyhow::Result<()> {
2222 let (_shape, root) = ctx.workspace_shape().ok_or_else(|| {
2223 workspace_not_initialised_error(
2224 "not inside a Memstead workspace (no `.memstead/workspace.toml` in any ancestor)",
2225 )
2226 })?;
2227
2228 let binding_id = args.binding;
2229
2230 // Parse the exclusions payload up front — a malformed `--exclusions` refuses
2231 // cheaply (before loading configs) with a typed code.
2232 let exclusions: std::collections::BTreeMap<String, String> =
2233 serde_json::from_str(&args.exclusions).map_err(|e| {
2234 CliError::new(
2235 ExitKind::Validation,
2236 "PROJECTION_INVALID_EXCLUSIONS",
2237 format!(
2238 "--exclusions must be a JSON object mapping in-scope artifact id → \
2239 rationale string: {e}"
2240 ),
2241 )
2242 .with_details(json!({ "error": e.to_string() }))
2243 })?;
2244
2245 // Find the binding by canonical id in the v1 store.
2246 let configs = load_pipeline_configs(&root).map_err(|e| {
2247 CliError::new(
2248 ExitKind::Generic,
2249 "PROJECTION_EXCLUDE_FAILED",
2250 format!("could not load pipeline config: {e}"),
2251 )
2252 .with_details(json!({ "error": e.to_string() }))
2253 })?;
2254 let record = configs
2255 .bindings
2256 .iter()
2257 .find(|r| format!("{}/{}", r.mem, r.name) == binding_id)
2258 .ok_or_else(|| binding_miss_error(&configs, &binding_id))?;
2259
2260 let resolved = resolve_binding_run(&binding_id, &record.config)
2261 .map_err(|e| map_resolve_err(&binding_id, e))?;
2262
2263 // The S(D) membership gate now spans every enumerable medium, including
2264 // graph — whose artifact set lives in the store, not on disk. So the
2265 // exclude path needs an engine exactly as verify does.
2266 let mut cli_engine = ctx.cli_engine_at(&root)?;
2267 let engine = cli_engine.base_mut();
2268
2269 let outcome = record_exclusions(engine, &root, &resolved, &exclusions)
2270 .map_err(|e| map_exclude_err(&binding_id, e))?;
2271
2272 if ctx.json {
2273 print_json(&json!({
2274 "binding": outcome.binding,
2275 "excluded": outcome.excluded,
2276 "added": outcome.added,
2277 // Each requested id and the canonical (workspace-relative) id
2278 // it resolved to: the spelling the ledger holds.
2279 "recorded": outcome
2280 .recorded
2281 .iter()
2282 .map(|(requested, canonical)| json!({
2283 "requested": requested,
2284 "canonical": canonical,
2285 }))
2286 .collect::<Vec<_>>(),
2287 }))?;
2288 } else {
2289 let mut body = format!(
2290 "# Projection exclude\n\nBinding `{}`: {} artifact(s) newly excluded, \
2291 {} in the ledger.\n",
2292 outcome.binding, outcome.added, outcome.excluded
2293 );
2294 for (requested, canonical) in &outcome.recorded {
2295 if requested == canonical {
2296 body.push_str(&format!("\n- `{canonical}`"));
2297 } else {
2298 body.push_str(&format!("\n- `{canonical}` (from `{requested}`)"));
2299 }
2300 }
2301 body.push('\n');
2302 print_markdown(&body);
2303 }
2304 Ok(())
2305}
2306
2307/// Render a one-block human note for the full-enumeration scheduling decision
2308/// (D3), prepended to the verify report so the typed signal is never silent: a
2309/// scheduled full walk that fired, a not-yet-due countdown, disabled scheduling,
2310/// and — critically — any non-enumerable refusal. Empty for the quiet cases
2311/// keeps a rotating-sample run byte-clean.
2312fn render_full_resync_note(decision: &FullResyncDecision) -> String {
2313 match decision {
2314 FullResyncDecision::Disabled => String::new(),
2315 FullResyncDecision::NotDue { .. } => String::new(),
2316 // An explicit full measurement (`--full`): every facet walked in
2317 // full, scheduler bypassed, cap unlimited — stated up front so the
2318 // report below reads as computed, not sampled.
2319 FullResyncDecision::Forced { walked_facets } => {
2320 let facets = if walked_facets.is_empty() {
2321 "(no primary facets)".to_string()
2322 } else {
2323 walked_facets.join(", ")
2324 };
2325 format!(
2326 "> **Full measurement (`--full`)** — full-enumeration walk over: {facets}. \
2327 Sampling scheduler bypassed; adjudication cap unlimited. Coverage and \
2328 accuracy figures below are computed over the whole source, not sampled.\n\n"
2329 )
2330 }
2331 FullResyncDecision::Due {
2332 walked_facets,
2333 refused,
2334 ..
2335 } => {
2336 let mut s = String::from("> **Scheduled full resync (D3)** — ");
2337 if walked_facets.is_empty() {
2338 s.push_str("no enumerable facet to walk this run.");
2339 } else {
2340 s.push_str(&format!(
2341 "full-enumeration coverage walk fired for: {}.",
2342 walked_facets.join(", ")
2343 ));
2344 }
2345 for r in refused {
2346 s.push_str(&format!(
2347 "\n> **Refused (cannot fully walk):** `{}` ({}) — {}",
2348 r.facet, r.medium_type, r.reason
2349 ));
2350 }
2351 s.push_str("\n\n");
2352 s
2353 }
2354 }
2355}
2356
2357/// `projection verify <binding>` — measure fidelity and record durable findings
2358/// (group A). Read-only on the destination mem's *entities*. A completed run
2359/// always records two things: the findings store, which is the verify surface's
2360/// own state outside the mem, and the prepared-hash backfill onto hash-less
2361/// anchors, which is measurement machinery (withhold it and an anchor never
2362/// leaves `recheck`, so drift stops being adjudicated). Its one FRESHNESS write,
2363/// the `#verified` baseline through the engine's sync-state writer, rides
2364/// `--advance` and is off by default, so a caller that verifies in order to read
2365/// leaves the mem's config byte-identical. An aborted or failed run makes none
2366/// of the three; in particular it never advances the baseline token.
2367fn verify(ctx: &CliContext, args: VerifyArgs) -> anyhow::Result<()> {
2368 let (_shape, root) = ctx.workspace_shape().ok_or_else(|| {
2369 workspace_not_initialised_error(
2370 "not inside a Memstead workspace (no `.memstead/workspace.toml` in any ancestor)",
2371 )
2372 })?;
2373
2374 let binding_id = args.binding;
2375
2376 let configs = load_pipeline_configs(&root).map_err(|e| {
2377 CliError::new(
2378 ExitKind::Generic,
2379 "PROJECTION_VERIFY_FAILED",
2380 format!("could not load pipeline config: {e}"),
2381 )
2382 .with_details(json!({ "error": e.to_string() }))
2383 })?;
2384 let record = configs
2385 .bindings
2386 .iter()
2387 .find(|r| format!("{}/{}", r.mem, r.name) == binding_id)
2388 .ok_or_else(|| binding_miss_error(&configs, &binding_id))?;
2389
2390 let resolved = resolve_binding_run(&binding_id, &record.config)
2391 .map_err(|e| map_resolve_err(&binding_id, e))?;
2392
2393 // The measurement pass takes a shared engine borrow (A5 — structurally
2394 // incapable of a mem mutation); the mutable binding exists only for the
2395 // completed-run baseline write below.
2396 let mut cli_engine = ctx.cli_engine_at(&root)?;
2397 let engine = cli_engine.base_mut();
2398
2399 // A quarantined destination serves NO entities, so a pass over it reads
2400 // every artifact as uncovered and every anchor as absent: the same
2401 // fiction the sidecar refusal below exists to prevent, one cause over.
2402 // Filed 2026-09-02 after a tick under a binary missing a schema version
2403 // loaded `engine` quarantined, then recorded 583 bogus uncovered
2404 // findings and a baseline, and pruned 181 authored dispositions.
2405 //
2406 // Placed HERE deliberately: above the measurement and above all three
2407 // stores it would otherwise touch (the findings store, the `#verified`
2408 // token, and the advance file whose exclusions `reconcile_exclusions`
2409 // prunes when the sources appear to hold nothing). A refusal further
2410 // down would leave whichever store sits above it already dirtied.
2411 if let Some(q) = engine.quarantine_reason(&resolved.destination_mem) {
2412 return Err(CliError::new(
2413 ExitKind::Validation,
2414 "MEM_QUARANTINED",
2415 format!(
2416 "verify refused for `{binding_id}`: the destination mem `{}` is \
2417quarantined ({}) — it serves no entities, so every artifact would read as \
2418uncovered and every anchor as absent, which is a measurement this run cannot \
2419honestly make. Repair the mem, then re-run",
2420 resolved.destination_mem, q.reason_message
2421 ),
2422 )
2423 .with_details(json!({
2424 "binding": binding_id,
2425 "mem": resolved.destination_mem,
2426 // The same vocabulary every other surface reports for this
2427 // condition, not a new one: code plus the boot reason.
2428 "reason_code": q.reason_code,
2429 "reason": q.reason_message,
2430 }))
2431 .into());
2432 }
2433
2434 // A malformed anchors sidecar reads as "no anchors", which a fidelity
2435 // pass would faithfully report as every artifact uncovered — findings,
2436 // and under `--fail-on-findings` a red build blaming the mem for a file
2437 // the engine could not parse. Refuse instead: the measurement cannot be
2438 // trusted, so this is an operational failure with its own code, never a
2439 // findings exit. Same posture the binding store takes when its own
2440 // record fails to load.
2441 if let Some(err) = engine.anchors_sidecar_error(&resolved.destination_mem) {
2442 return Err(CliError::new(
2443 ExitKind::Validation,
2444 "ANCHORS_SIDECAR_UNREADABLE",
2445 format!(
2446 "verify refused for `{binding_id}`: the anchors sidecar for mem `{}` \
2447does not parse ({err}) — every anchor would read as absent and every artifact \
2448as uncovered, which is a measurement this run cannot honestly make. Repair or \
2449remove the sidecar and re-run",
2450 resolved.destination_mem
2451 ),
2452 )
2453 .with_details(json!({
2454 "binding": binding_id,
2455 "mem": resolved.destination_mem,
2456 // The same key every other surface uses for this condition.
2457 "reason": err,
2458 }))
2459 .into());
2460 }
2461
2462 let run = if args.full {
2463 verify_binding_full
2464 } else {
2465 verify_binding
2466 };
2467 let outcome = run(engine, &root, &record.config, &resolved).map_err(|e| match &e {
2468 // A vanished/unmounted source is a typed refusal, not a failed
2469 // measurement: nothing was observed, no findings were recorded,
2470 // and the `#verified` baseline is deliberately left untouched
2471 // (a transient unmount must never clobber real recorded state).
2472 FindingsError::SourceUnreachable { source_name, path } => CliError::new(
2473 ExitKind::Validation,
2474 "SOURCE_UNREACHABLE",
2475 format!(
2476 "verify refused for `{binding_id}`: source '{source_name}' resolves to \
2477 `{path}`, which cannot be read (absent, or present but not \
2478 enumerable) — restore or remount the source (or \
2479 repoint its pointer); the recorded `#verified` baseline was left \
2480 untouched"
2481 ),
2482 )
2483 .with_details(json!({
2484 "binding": binding_id,
2485 "source": source_name,
2486 "path": path,
2487 })),
2488 // `--full` over a non-enumerable medium: the existing typed
2489 // capability refusal — a full measurement promises complete
2490 // figures, so the run refuses instead of rendering a report
2491 // with fabricated completeness. Nothing was observed or
2492 // recorded.
2493 FindingsError::FullWalkNonEnumerable(refusal) => CliError::new(
2494 ExitKind::Validation,
2495 "PROJECTION_CAPABILITY_UNSUPPORTED",
2496 format!("verify --full refused for `{binding_id}`: {e}"),
2497 )
2498 .with_details(json!({
2499 "binding": binding_id,
2500 "facet": refusal.facet,
2501 "medium_type": refusal.medium_type,
2502 "reason": refusal.reason,
2503 })),
2504 _ => CliError::new(
2505 ExitKind::Generic,
2506 "PROJECTION_VERIFY_FAILED",
2507 format!("verify failed for `{binding_id}`: {e}"),
2508 )
2509 .with_details(json!({ "binding": binding_id, "error": e.to_string() })),
2510 })?;
2511
2512 // The run completed — record its prepared-hash backfill: every hash the
2513 // pass observed for a hash-less hash-bearing anchor lands on that anchor
2514 // in the engine-owned anchors sidecar (measurement bookkeeping — no
2515 // entity content is touched). Before the report, so the rendered
2516 // anchor-resolution figures reflect the recorded hashes. Idempotent: a
2517 // pass over fully-backfilled anchors observes an empty worklist.
2518 // Bookkeeping-write failures are reported AFTER the report, never
2519 // instead of it. Both of these say "verify completed and findings were
2520 // recorded" — a run that completed owes the caller its measurement, and
2521 // returning here would hand a CI job a red build with nothing to read.
2522 // Same render-then-fail ordering the findings gate uses, extended to the
2523 // paths that can fail between the measurement and the render.
2524 // The backfill is NOT gated on `--advance`, and the distinction is the
2525 // point: it is measurement machinery, not a freshness baseline. A
2526 // hash-less anchor that never receives its observed hash stays in
2527 // `recheck` forever, so drift is never adjudicated deterministically and
2528 // the verdict itself changes. Gating it was tried and reverted (C6,
2529 // 2026-09-03) on that evidence: seven projection tests moved from
2530 // `drifted` to `clean` with the backfill withheld. What `--advance`
2531 // gates is the `#verified` token, which is the freshness claim a later
2532 // run reads, and which is what the entry was filed on.
2533 let backfill_result = record_anchor_hash_backfill(
2534 engine,
2535 &resolved.destination_mem,
2536 &outcome,
2537 Some("projection verify: prepared-hash backfill onto hash-less anchors"),
2538 );
2539 let hashes_backfilled = *backfill_result.as_ref().unwrap_or(&0);
2540
2541 // Assemble + render the tier-1 fidelity report (group B) over the findings
2542 // the pass just recorded. Read-only — no destination-mem mutation.
2543 let budget = args.budget.unwrap_or(DEFAULT_REPORT_BUDGET);
2544 let report = compute_fidelity_report(engine, &root, &record.config, &resolved, &outcome.key);
2545 let rendered = render_fidelity_report(&report, budget, &args.include);
2546
2547 // The run completed — record its `#verified` baseline per observed facet
2548 // head through the engine's sync-state writer (the backlog-prescribed
2549 // writer; a failed run returned above and never reaches this).
2550 let baseline_result = if args.advance {
2551 record_verified_baseline(
2552 engine,
2553 &resolved.destination_mem,
2554 &outcome,
2555 Some("projection verify --advance: completed-run #verified baseline"),
2556 )
2557 } else {
2558 Ok(Vec::new())
2559 };
2560 let verified_baseline = baseline_result.as_ref().cloned().unwrap_or_default();
2561
2562 // The rollup is derived from the assembled report, so the headline a
2563 // human reads, the `rollup` block a consumer parses, and the gate exit
2564 // code below are all the same judgement — they cannot disagree.
2565 let rollup = report.rollup();
2566
2567 if ctx.json {
2568 print_json(&json!({
2569 // Version marker in the house style (`memstead-export/v1`,
2570 // `workspace-dump/v1`): consumers assert it before parsing, so a
2571 // future shape change fails loudly instead of misparsing. This
2572 // payload is external contract — see the verify-in-CI guide.
2573 "format": JSON_VERIFY_FORMAT,
2574 // The coverage rule (memstead_base::ops::coverage): the
2575 // axes the rollup verdict answers for; the per-run
2576 // blind-spots inside `rollup` are the refinement.
2577 "verdict_coverage": crate::coverage::PROJECTION_VERIFY
2578 .axis_coverage()
2579 .expect("projection verify is a verdict surface")
2580 .wire_line(),
2581 "rollup": rollup,
2582 "binding": outcome.binding,
2583 "key": {
2584 "binding_hash": outcome.key.binding_hash,
2585 "source_head": outcome.key.source_head,
2586 },
2587 "recorded": outcome.recorded,
2588 "superseded": outcome.superseded,
2589 "backlog": outcome.backlog,
2590 // The tier-3 full-enumeration scheduling decision (D3) — surfaced
2591 // (never a silent skip): whether a scheduled full walk fired, is not
2592 // yet due, is disabled, and any typed non-enumerable refusals.
2593 "full_resync": outcome.full_resync,
2594 // Whether this run advanced the `#verified` freshness baseline
2595 // (`--advance`). False is the default, so an empty
2596 // `verified_baseline` below means "not asked for", not "nothing
2597 // to record". The backfill count is unaffected by this flag.
2598 "advanced": args.advance,
2599 // The completed run's `#verified` baseline keys, written through
2600 // the engine's sync-state writer. Empty without `--advance`.
2601 "verified_baseline": verified_baseline,
2602 // How many hash-less hash-bearing anchors gained a recorded
2603 // prepared-content hash this run (the completed-run backfill
2604 // write into the engine-owned anchors sidecar). 0 once every
2605 // anchor carries its hash — the backfill is idempotent.
2606 "hash_backfilled": hashes_backfilled,
2607 "report": report,
2608 "report_mode": rendered.mode,
2609 "report_markdown": rendered.markdown,
2610 }))?;
2611 } else {
2612 // The rendered report IS the stdout content (agent-consumable brief);
2613 // prepend the scheduled full-walk decision so D3's typed signal (a full
2614 // sweep, or a non-enumerable refusal) is never silent in human mode,
2615 // and append the recorded `#verified` baseline so the completed-run
2616 // write is visible.
2617 let baseline_note = if !args.advance {
2618 // Never silent: a reader must be able to tell "wrote nothing
2619 // because that is the default" from "had nothing to write".
2620 "\n> **Baseline not advanced** \u{2014} this run measured without moving the \
2621 `#verified` token, so the mem's config is untouched and the next run still \
2622 sees the binding as unverified. Pass `--advance` to record it.\n"
2623 .to_string()
2624 } else if verified_baseline.is_empty() {
2625 String::new()
2626 } else {
2627 format!(
2628 "\n> **Verified baseline recorded** — {}\n",
2629 verified_baseline
2630 .iter()
2631 .map(|k| format!("`{k}`"))
2632 .collect::<Vec<_>>()
2633 .join(", ")
2634 )
2635 };
2636 let backfill_note = if hashes_backfilled == 0 {
2637 String::new()
2638 } else {
2639 format!(
2640 "\n> **Prepared-hash backfill recorded** — {hashes_backfilled} hash-less \
2641 anchor(s) now carry their observed prepared-content hash; subsequent \
2642 verifies adjudicate them deterministically.\n"
2643 )
2644 };
2645 // The coverage rule: the axes the rollup verdict answers for,
2646 // in the human rendering too (the JSON envelope stamps above;
2647 // memstead_base::ops::coverage).
2648 let coverage_note = crate::coverage::PROJECTION_VERIFY
2649 .axis_coverage()
2650 .map(|cov| format!("\n**Verdict coverage:** {}\n", cov.wire_line()))
2651 .unwrap_or_default();
2652 print_markdown(&format!(
2653 "{}{}{}{}{}",
2654 render_full_resync_note(&outcome.full_resync),
2655 rendered.markdown,
2656 coverage_note,
2657 backfill_note,
2658 baseline_note
2659 ));
2660 }
2661
2662 // --- Bookkeeping-write failures, now that the report has been rendered ---
2663 // A failed write is an operational failure with its own code, never the
2664 // findings exit: the measurement's own answer is already on stdout above,
2665 // and this says the run could not finish recording it.
2666 if let Err(e) = backfill_result {
2667 return Err(CliError::new(
2668 ExitKind::Generic,
2669 "PROJECTION_VERIFY_BACKFILL_FAILED",
2670 format!(
2671 "verify completed and findings were recorded for `{binding_id}` (the report is \
2672above), but recording the prepared-hash backfill onto the anchors sidecar failed: {e}"
2673 ),
2674 )
2675 .with_details(json!({ "binding": binding_id, "error": e.to_string() }))
2676 .into());
2677 }
2678 if let Err(e) = baseline_result {
2679 return Err(CliError::new(
2680 ExitKind::Generic,
2681 "PROJECTION_VERIFY_BASELINE_FAILED",
2682 format!(
2683 "verify completed and findings were recorded for `{binding_id}` (the report is \
2684above), but writing the `#verified` baseline failed: {e} — the next run will treat this \
2685binding as never verified"
2686 ),
2687 )
2688 .with_details(json!({ "binding": binding_id, "error": e.to_string() }))
2689 .into());
2690 }
2691
2692 // --- CI gate (opt-in) ---
2693 // Report first, then fail: `main` prints the error envelope and nothing
2694 // else, so a findings exit that returned before this point would hand a
2695 // CI job a red build with no report to read. Same ordering `health
2696 // --strict` uses, with the ambiguity that one has removed — this code is
2697 // dedicated, so a job can tell "the mem drifted" from "the engine failed
2698 // to boot".
2699 if args.fail_on_findings && rollup.findings_total > 0 {
2700 return Err(CliError::new(
2701 ExitKind::Findings,
2702 "PROJECTION_VERIFY_FINDINGS",
2703 format!(
2704 "verify completed for `{binding_id}` and recorded {} finding(s) — {}",
2705 rollup.findings_total, rollup.because
2706 ),
2707 )
2708 .with_details(json!({
2709 "binding": binding_id,
2710 "verdict": rollup.verdict.wire(),
2711 "findings_total": rollup.findings_total,
2712 "findings_by_class": report.findings_by_class,
2713 "actions": rollup.actions,
2714 }))
2715 .into());
2716 }
2717 // --- Inconclusive gate (opt-in) ---
2718 // Same report-first ordering; evaluated AFTER the findings gate, so
2719 // a findings run with both flags exits with the findings code (a
2720 // substantive result outranks a blindness report). Without this
2721 // flag an inconclusive run keeps its long-standing exit 0 — the
2722 // gate exists because that 0 is indistinguishable from a
2723 // substantive clean pass to a code-only consumer.
2724 if args.fail_on_inconclusive
2725 && rollup.verdict == memstead_base::ingest::report::RollupVerdict::Inconclusive
2726 {
2727 return Err(CliError::new(
2728 ExitKind::Findings,
2729 "PROJECTION_VERIFY_INCONCLUSIVE",
2730 format!(
2731 "verify completed for `{binding_id}` but the measurement was blind — verdict \
2732 inconclusive: {}",
2733 rollup.because
2734 ),
2735 )
2736 .with_details(json!({
2737 "binding": binding_id,
2738 "verdict": rollup.verdict.wire(),
2739 "blind_spots": rollup.blind_spots,
2740 "actions": rollup.actions,
2741 }))
2742 .into());
2743 }
2744 Ok(())
2745}