Skip to main content

harn_vm/orchestration/
release_fixture.rs

1//! Release-harness fixture ingest.
2//!
3//! `release_harn.harn` (in `harn-bump-fleet`) emits a
4//! `release_harn.crystallization_input.v1` fixture bundle for every run:
5//!
6//! ```text
7//! crystallization-input/
8//!   manifest.json                # release identity + file map
9//!   release-run.json             # full release-harness payload
10//!   deterministic-events.jsonl   # harness facts/findings/step records
11//!   agent-events.jsonl           # model review/changelog/recovery events
12//!   tool-observations.jsonl      # shell/read/tool observations
13//!   README.md                    # human-readable description
14//! ```
15//!
16//! This module ingests one of those bundles and synthesizes a single
17//! crystallization candidate (no repeated-sequence mining required —
18//! the trace IS the workflow). Output is a [`CrystallizationArtifacts`]
19//! that downstream tooling (`build_crystallization_bundle`,
20//! `validate_crystallization_bundle`, `shadow_replay_bundle`) consumes
21//! unchanged.
22//!
23//! See `harn-bump-fleet/release_harn.harn` for the upstream emitter and
24//! `docs/src/workflow-crystallization.md` for the steel-thread story.
25
26use std::collections::{BTreeMap, BTreeSet};
27use std::path::Path;
28
29use serde::{Deserialize, Serialize};
30use serde_json::{json, Value as JsonValue};
31use sha2::{Digest, Sha256};
32
33use super::crystallize::{
34    synthesize_candidate_from_trace, CrystallizationAction, CrystallizationApproval,
35    CrystallizationArtifacts, CrystallizationSideEffect, CrystallizationTrace, CrystallizeOptions,
36    RecoveryFeedbackSummary, SegmentSummary, WorkflowCandidateParameter,
37};
38use crate::value::VmError;
39
40/// Schema marker emitted by `release_harn.harn` on every event/manifest
41/// row. The trailing `.vN` carries the version. Importers MUST refuse
42/// fixtures with any other schema marker — bumping the suffix is how
43/// `release_harn.harn` signals a backwards-incompatible fixture shape
44/// change.
45pub const RELEASE_FIXTURE_SCHEMA: &str = "release_harn.crystallization_input.v1";
46
47const MANIFEST_FILE: &str = "manifest.json";
48const DETERMINISTIC_EVENTS_FILE: &str = "deterministic-events.jsonl";
49const AGENT_EVENTS_FILE: &str = "agent-events.jsonl";
50const TOOL_OBSERVATIONS_FILE: &str = "tool-observations.jsonl";
51
52/// Decoded release-fixture manifest. We keep this struct intentionally
53/// small — only the fields the ingester actually depends on — and
54/// preserve unknown fields verbatim in [`ReleaseFixtureManifest::raw`]
55/// so future schema additions don't lose information when the fixture
56/// is round-tripped.
57#[derive(Clone, Debug, Default, Serialize, Deserialize)]
58pub struct ReleaseFixtureManifest {
59    pub schema_version: String,
60    pub kind: String,
61    pub generated_by: String,
62    pub generated_at: String,
63    pub run_id: String,
64    pub source_issue: Option<String>,
65    pub consumer_issue: Option<String>,
66    pub release: ReleaseFixtureIdentity,
67    /// The full untouched manifest JSON. Useful for downstream
68    /// consumers and for round-tripping unknown fields without losing
69    /// them.
70    #[serde(skip_serializing, skip_deserializing)]
71    pub raw: Option<JsonValue>,
72}
73
74#[derive(Clone, Debug, Default, Serialize, Deserialize)]
75pub struct ReleaseFixtureIdentity {
76    pub repo: String,
77    pub mode: String,
78    #[serde(default)]
79    pub mock: bool,
80    pub current_version: String,
81    pub next_version: String,
82    pub starting_branch: String,
83    pub target_release_branch: String,
84    pub base_branch: String,
85    pub latest_tag: String,
86}
87
88/// One row from one of the JSONL fixtures (deterministic-events,
89/// agent-events, tool-observations).
90#[derive(Clone, Debug, Default, Serialize, Deserialize)]
91pub struct ReleaseFixtureEvent {
92    pub schema_version: String,
93    pub event_type: String,
94    pub source: String,
95    #[serde(default)]
96    pub timestamp: Option<String>,
97    #[serde(default)]
98    pub data: JsonValue,
99}
100
101/// All decoded data from one ingested fixture bundle. Kept distinct
102/// from the eventual [`CrystallizationTrace`] so the importer can run
103/// secondary analysis (segment summary, recovery summary) over the
104/// original event stream.
105#[derive(Clone, Debug, Default)]
106pub struct ReleaseFixture {
107    pub manifest: ReleaseFixtureManifest,
108    pub deterministic_events: Vec<ReleaseFixtureEvent>,
109    pub agent_events: Vec<ReleaseFixtureEvent>,
110    pub tool_observations: Vec<ReleaseFixtureEvent>,
111}
112
113/// Read a release-fixture bundle directory from disk. The manifest is
114/// required; missing JSONL streams are treated as empty (so a fixture
115/// from a release with no agent review still ingests cleanly).
116pub fn load_release_fixture(dir: &Path) -> Result<ReleaseFixture, VmError> {
117    let manifest_path = dir.join(MANIFEST_FILE);
118    let manifest_bytes = std::fs::read(&manifest_path).map_err(|error| {
119        VmError::Runtime(format!(
120            "failed to read release fixture manifest {}: {error}",
121            manifest_path.display()
122        ))
123    })?;
124    let manifest_json: JsonValue = serde_json::from_slice(&manifest_bytes).map_err(|error| {
125        VmError::Runtime(format!(
126            "failed to parse release fixture manifest {}: {error}",
127            manifest_path.display()
128        ))
129    })?;
130    let mut manifest: ReleaseFixtureManifest = serde_json::from_value(manifest_json.clone())
131        .map_err(|error| {
132            VmError::Runtime(format!(
133                "failed to decode release fixture manifest {}: {error}",
134                manifest_path.display()
135            ))
136        })?;
137    manifest.raw = Some(manifest_json);
138
139    if manifest.schema_version != RELEASE_FIXTURE_SCHEMA {
140        return Err(VmError::Runtime(format!(
141            "release fixture manifest {} has unrecognized schema_version {:?} (expected {})",
142            manifest_path.display(),
143            manifest.schema_version,
144            RELEASE_FIXTURE_SCHEMA
145        )));
146    }
147
148    let deterministic_events = load_optional_jsonl(&dir.join(DETERMINISTIC_EVENTS_FILE))?;
149    let agent_events = load_optional_jsonl(&dir.join(AGENT_EVENTS_FILE))?;
150    let tool_observations = load_optional_jsonl(&dir.join(TOOL_OBSERVATIONS_FILE))?;
151
152    Ok(ReleaseFixture {
153        manifest,
154        deterministic_events,
155        agent_events,
156        tool_observations,
157    })
158}
159
160fn load_optional_jsonl(path: &Path) -> Result<Vec<ReleaseFixtureEvent>, VmError> {
161    if !path.exists() {
162        return Ok(Vec::new());
163    }
164    let content = std::fs::read_to_string(path).map_err(|error| {
165        VmError::Runtime(format!(
166            "failed to read release fixture stream {}: {error}",
167            path.display()
168        ))
169    })?;
170    let mut events = Vec::new();
171    for (line_idx, line) in content.lines().enumerate() {
172        let trimmed = line.trim();
173        if trimmed.is_empty() {
174            continue;
175        }
176        let event: ReleaseFixtureEvent = serde_json::from_str(trimmed).map_err(|error| {
177            VmError::Runtime(format!(
178                "failed to decode release fixture event {} line {}: {error}",
179                path.display(),
180                line_idx + 1
181            ))
182        })?;
183        if event.schema_version != RELEASE_FIXTURE_SCHEMA {
184            return Err(VmError::Runtime(format!(
185                "release fixture event {} line {} has unrecognized schema {:?} (expected {})",
186                path.display(),
187                line_idx + 1,
188                event.schema_version,
189                RELEASE_FIXTURE_SCHEMA
190            )));
191        }
192        events.push(event);
193    }
194    Ok(events)
195}
196
197/// Convert a release fixture into a single [`CrystallizationTrace`].
198/// Each release-step / tool-observation / agent-attempt /
199/// recovery-advice event becomes one [`CrystallizationAction`], in
200/// timestamp order. Side effects are inferred from
201/// `release_step_recorded` events; capabilities and approval boundaries
202/// are derived from event type + tool kind.
203pub fn release_fixture_to_trace(fixture: &ReleaseFixture) -> CrystallizationTrace {
204    let release = &fixture.manifest.release;
205    let mut metadata: BTreeMap<String, JsonValue> = BTreeMap::new();
206    metadata.insert("release.repo".to_string(), json!(release.repo));
207    metadata.insert("release.mode".to_string(), json!(release.mode));
208    metadata.insert("release.mock".to_string(), json!(release.mock));
209    metadata.insert(
210        "release.current_version".to_string(),
211        json!(release.current_version),
212    );
213    metadata.insert(
214        "release.next_version".to_string(),
215        json!(release.next_version),
216    );
217    metadata.insert(
218        "release.starting_branch".to_string(),
219        json!(release.starting_branch),
220    );
221    metadata.insert(
222        "release.target_release_branch".to_string(),
223        json!(release.target_release_branch),
224    );
225    metadata.insert(
226        "release.base_branch".to_string(),
227        json!(release.base_branch),
228    );
229    metadata.insert("release.latest_tag".to_string(), json!(release.latest_tag));
230    metadata.insert("fixture.run_id".to_string(), json!(fixture.manifest.run_id));
231    metadata.insert(
232        "fixture.generated_by".to_string(),
233        json!(fixture.manifest.generated_by),
234    );
235
236    let mut actions = Vec::new();
237    for event in &fixture.deterministic_events {
238        if let Some(action) = deterministic_event_to_action(event, release) {
239            actions.push(action);
240        }
241    }
242    for event in &fixture.tool_observations {
243        if let Some(action) = tool_observation_to_action(event, release) {
244            actions.push(action);
245        }
246    }
247    for event in &fixture.agent_events {
248        if let Some(action) = agent_event_to_action(event, release) {
249            actions.push(action);
250        }
251    }
252
253    actions.sort_by(|left, right| left.timestamp.cmp(&right.timestamp));
254    // Re-key after sort so action ids stay deterministic and ordered.
255    for (idx, action) in actions.iter_mut().enumerate() {
256        if action.id.trim().is_empty() {
257            action.id = format!("event_{}", idx + 1);
258        }
259    }
260
261    let trace_id = if fixture.manifest.run_id.trim().is_empty() {
262        format!(
263            "release_fixture_{}",
264            hash_short(fixture.manifest.generated_at.as_bytes())
265        )
266    } else {
267        format!("release_fixture_{}", fixture.manifest.run_id)
268    };
269
270    CrystallizationTrace {
271        version: 1,
272        id: trace_id,
273        source: Some(format!(
274            "release_harn.harn run {} ({} -> {})",
275            fixture.manifest.run_id, release.current_version, release.next_version
276        )),
277        source_hash: None,
278        workflow_id: Some("release_harn".to_string()),
279        started_at: actions.first().and_then(|action| action.timestamp.clone()),
280        finished_at: actions.last().and_then(|action| action.timestamp.clone()),
281        flow: None,
282        actions,
283        replay_run: None,
284        replay_allowlist: Vec::new(),
285        usage: Default::default(),
286        metadata,
287    }
288}
289
290fn deterministic_event_to_action(
291    event: &ReleaseFixtureEvent,
292    release: &ReleaseFixtureIdentity,
293) -> Option<CrystallizationAction> {
294    let mut action = base_action_from_event(event);
295    let data = &event.data;
296    match event.event_type.as_str() {
297        "release_analysis" => {
298            action.kind = "release_analysis".to_string();
299            action.name = format!(
300                "analyze_release[{}->{}]",
301                data.get("current_version")
302                    .and_then(JsonValue::as_str)
303                    .unwrap_or(&release.current_version),
304                data.get("next_version")
305                    .and_then(JsonValue::as_str)
306                    .unwrap_or(&release.next_version),
307            );
308            action.parameters = release_parameters(release);
309            action.deterministic = Some(true);
310            action.fuzzy = Some(false);
311            action.capabilities = vec!["git.read".to_string(), "fs.read".to_string()];
312            action.inputs = data.clone();
313        }
314        "changelog_audit_inputs" => {
315            action.kind = "changelog_audit".to_string();
316            action.name = "collect_changelog_audit_inputs".to_string();
317            action.parameters = release_parameters(release);
318            action.deterministic = Some(true);
319            action.fuzzy = Some(false);
320            action.capabilities = vec!["fs.read".to_string()];
321            action.inputs = data.clone();
322        }
323        "deterministic_finding" => {
324            let finding = data
325                .get("finding")
326                .and_then(JsonValue::as_str)
327                .unwrap_or("finding");
328            action.kind = "release_finding".to_string();
329            action.name = format!("finding[{}]", short_summary(finding, 32));
330            action.deterministic = Some(true);
331            action.fuzzy = Some(false);
332            action.observed_output = Some(json!({"finding": finding}));
333        }
334        "release_step_recorded" => {
335            let step_name = data
336                .get("name")
337                .and_then(JsonValue::as_str)
338                .unwrap_or("step")
339                .to_string();
340            let command = data
341                .get("command")
342                .and_then(JsonValue::as_str)
343                .unwrap_or("")
344                .to_string();
345            let success = data
346                .get("success")
347                .and_then(JsonValue::as_bool)
348                .unwrap_or(true);
349            let status = data.get("status").cloned().unwrap_or(json!(0));
350            let classification = data
351                .get("classification")
352                .and_then(JsonValue::as_str)
353                .unwrap_or("");
354            let recovery_hint = data
355                .get("recovery_hint")
356                .and_then(JsonValue::as_str)
357                .unwrap_or("");
358            action.kind = if success {
359                "release_step".to_string()
360            } else {
361                "shell_failure".to_string()
362            };
363            action.name = sanitize_step_name(&step_name);
364            action.deterministic = Some(true);
365            action.fuzzy = Some(false);
366            action.parameters = BTreeMap::from([
367                ("step".to_string(), json!(step_name)),
368                ("command".to_string(), json!(command)),
369            ]);
370            action.inputs = json!({"command": command});
371            action.observed_output = Some(json!({"status": status, "success": success}));
372            if !success {
373                action.side_effects.push(CrystallizationSideEffect {
374                    kind: "shell_failure".to_string(),
375                    target: step_name,
376                    capability: Some("shell.exec".to_string()),
377                    mutation: None,
378                    metadata: BTreeMap::from([(
379                        "classification".to_string(),
380                        json!(classification),
381                    )]),
382                });
383            }
384            action
385                .metadata
386                .insert("classification".to_string(), json!(classification));
387            action
388                .metadata
389                .insert("recovery_hint".to_string(), json!(recovery_hint));
390            action
391                .metadata
392                .insert("success".to_string(), json!(success));
393        }
394        _ => return None,
395    }
396    Some(action)
397}
398
399fn tool_observation_to_action(
400    event: &ReleaseFixtureEvent,
401    release: &ReleaseFixtureIdentity,
402) -> Option<CrystallizationAction> {
403    if event.event_type != "tool_observation" {
404        return None;
405    }
406    let mut action = base_action_from_event(event);
407    let data = &event.data;
408    let tool = data
409        .get("tool")
410        .and_then(JsonValue::as_str)
411        .unwrap_or("shell");
412    let command = data
413        .get("command")
414        .and_then(JsonValue::as_str)
415        .unwrap_or("");
416    let path = data.get("path").and_then(JsonValue::as_str);
417    let step_name = data.get("step_name").and_then(JsonValue::as_str);
418    let success = data
419        .get("success")
420        .and_then(JsonValue::as_bool)
421        .unwrap_or(true);
422
423    action.kind = format!("tool_observation:{tool}");
424    action.name = match (tool, step_name) {
425        (_, Some(name)) if !name.is_empty() => sanitize_step_name(name),
426        ("read_file", Some(_)) | ("read_file", None) => format!(
427            "read_file[{}]",
428            short_summary(path.unwrap_or("unknown"), 48)
429        ),
430        ("shell", _) => format!("shell[{}]", short_summary(command, 48)),
431        _ => format!("{tool}[{}]", short_summary(command, 48)),
432    };
433    action.deterministic = Some(event.source != "agent");
434    action.fuzzy = Some(false);
435    let _ = release; // release identity only enters via metadata above
436    action.inputs = data.clone();
437    action.observed_output = Some(json!({
438        "stdout": data.get("stdout").cloned().unwrap_or(json!("")),
439        "stderr": data.get("stderr").cloned().unwrap_or(json!("")),
440        "status": data.get("status").cloned().unwrap_or(json!(0)),
441        "success": success,
442    }));
443    if !success {
444        action.side_effects.push(CrystallizationSideEffect {
445            kind: "tool_failure".to_string(),
446            target: step_name.unwrap_or(tool).to_string(),
447            capability: Some("shell.exec".to_string()),
448            mutation: None,
449            metadata: BTreeMap::default(),
450        });
451    }
452    action.metadata.insert("tool".to_string(), json!(tool));
453    Some(action)
454}
455
456fn agent_event_to_action(
457    event: &ReleaseFixtureEvent,
458    release: &ReleaseFixtureIdentity,
459) -> Option<CrystallizationAction> {
460    let mut action = base_action_from_event(event);
461    let data = &event.data;
462    match event.event_type.as_str() {
463        "agent_review_attempt" => {
464            let name = data
465                .get("name")
466                .and_then(JsonValue::as_str)
467                .unwrap_or("review");
468            action.kind = "agent_review_attempt".to_string();
469            action.name = format!("agent_review[{}]", sanitize_step_name(name));
470            action.fuzzy = Some(true);
471            action.deterministic = Some(false);
472            action.cost.model_calls = 1;
473            action.parameters = release_parameters(release);
474            action.inputs = data.clone();
475            action.observed_output = data.get("text").cloned();
476            action.approval = Some(CrystallizationApproval {
477                prompt: Some(
478                    "Human reviewer must accept the agent-authored audit before promotion."
479                        .to_string(),
480                ),
481                approver: None,
482                required: true,
483                boundary: Some("release_audit_review".to_string()),
484            });
485        }
486        "agent_review_artifacts" => {
487            action.kind = "agent_review_artifacts".to_string();
488            action.name = "persist_agent_review_artifacts".to_string();
489            action.deterministic = Some(true);
490            action.fuzzy = Some(false);
491            action.observed_output = Some(data.clone());
492        }
493        "agent_recovery_advice" => {
494            let step_name = data
495                .get("step_name")
496                .and_then(JsonValue::as_str)
497                .unwrap_or("recovery");
498            action.kind = "agent_recovery_advice".to_string();
499            action.name = format!("recovery_advice[{}]", sanitize_step_name(step_name));
500            action.fuzzy = Some(true);
501            action.deterministic = Some(false);
502            action.cost.model_calls = 1;
503            action.observed_output = data.get("text").cloned();
504            action.parameters = BTreeMap::from([("failed_step".to_string(), json!(step_name))]);
505            action.approval = Some(CrystallizationApproval {
506                prompt: Some(
507                    "Human reviewer must validate recovery advice before re-running the failing step."
508                        .to_string(),
509                ),
510                approver: None,
511                required: true,
512                boundary: Some("recovery_review".to_string()),
513            });
514            action
515                .metadata
516                .insert("failed_step".to_string(), json!(step_name));
517        }
518        "agent_review_skipped" => {
519            action.kind = "agent_review_skipped".to_string();
520            action.name = "agent_review_skipped".to_string();
521            action.deterministic = Some(true);
522            action.fuzzy = Some(false);
523            action.observed_output = data.get("reason").cloned();
524        }
525        _ => return None,
526    }
527    Some(action)
528}
529
530fn base_action_from_event(event: &ReleaseFixtureEvent) -> CrystallizationAction {
531    CrystallizationAction {
532        id: stable_event_id(event),
533        timestamp: event.timestamp.clone(),
534        ..CrystallizationAction::default()
535    }
536}
537
538fn stable_event_id(event: &ReleaseFixtureEvent) -> String {
539    let mut hasher = Sha256::new();
540    hasher.update(event.event_type.as_bytes());
541    hasher.update([0]);
542    hasher.update(event.source.as_bytes());
543    hasher.update([0]);
544    if let Some(timestamp) = &event.timestamp {
545        hasher.update(timestamp.as_bytes());
546    }
547    hasher.update([0]);
548    hasher.update(event.data.to_string().as_bytes());
549    let hex = hex::encode(hasher.finalize());
550    format!(
551        "evt_{}_{}",
552        sanitize_step_name(&event.event_type),
553        hex.chars().take(12).collect::<String>()
554    )
555}
556
557fn release_parameters(release: &ReleaseFixtureIdentity) -> BTreeMap<String, JsonValue> {
558    let mut out = BTreeMap::new();
559    out.insert("repo".to_string(), json!(release.repo));
560    out.insert("base_branch".to_string(), json!(release.base_branch));
561    out.insert(
562        "current_version".to_string(),
563        json!(release.current_version),
564    );
565    out.insert("next_version".to_string(), json!(release.next_version));
566    out.insert(
567        "target_release_branch".to_string(),
568        json!(release.target_release_branch),
569    );
570    out
571}
572
573fn sanitize_step_name(raw: &str) -> String {
574    let cleaned: String = raw
575        .chars()
576        .map(|ch| {
577            if ch.is_ascii_alphanumeric() || ch == '_' || ch == '-' {
578                ch.to_ascii_lowercase()
579            } else {
580                '_'
581            }
582        })
583        .collect();
584    let trimmed = cleaned.trim_matches('_').to_string();
585    if trimmed.is_empty() {
586        "step".to_string()
587    } else {
588        trimmed
589    }
590}
591
592fn short_summary(raw: &str, max_chars: usize) -> String {
593    let cleaned = raw.replace(['\n', '\r', '\t'], " ");
594    let trimmed = cleaned.trim();
595    if trimmed.chars().count() <= max_chars {
596        trimmed.to_string()
597    } else {
598        let prefix: String = trimmed.chars().take(max_chars - 1).collect();
599        format!("{prefix}…")
600    }
601}
602
603fn hash_short(bytes: &[u8]) -> String {
604    let mut hasher = Sha256::new();
605    hasher.update(bytes);
606    hex::encode(hasher.finalize()).chars().take(12).collect()
607}
608
609/// Build a [`SegmentSummary`] from an ingested fixture. Splits steps
610/// into "safe to automate" (deterministic events with no shell
611/// failures) vs. "requires human review" (agent events, shell
612/// failures, recovery advice).
613pub fn build_segment_summary(fixture: &ReleaseFixture) -> SegmentSummary {
614    let mut safe = Vec::new();
615    let mut review = Vec::new();
616    for event in &fixture.deterministic_events {
617        match event.event_type.as_str() {
618            "release_analysis" | "changelog_audit_inputs" => safe.push(event.event_type.clone()),
619            "release_step_recorded" => {
620                let name = event
621                    .data
622                    .get("name")
623                    .and_then(JsonValue::as_str)
624                    .unwrap_or("release_step")
625                    .to_string();
626                let success = event
627                    .data
628                    .get("success")
629                    .and_then(JsonValue::as_bool)
630                    .unwrap_or(true);
631                if success {
632                    safe.push(format!("step:{name}"));
633                } else {
634                    review.push(format!(
635                        "failed step:{name} (deterministic recovery required before re-run)"
636                    ));
637                }
638            }
639            "deterministic_finding" => {
640                let finding = event
641                    .data
642                    .get("finding")
643                    .and_then(JsonValue::as_str)
644                    .unwrap_or("finding");
645                review.push(format!("finding requires reviewer attention: {finding}"));
646            }
647            _ => {}
648        }
649    }
650    for event in &fixture.agent_events {
651        match event.event_type.as_str() {
652            "agent_review_attempt" => {
653                let name = event
654                    .data
655                    .get("name")
656                    .and_then(JsonValue::as_str)
657                    .unwrap_or("agent_review");
658                review.push(format!(
659                    "agent review attempt:{name} (model-authored, advisory only)"
660                ));
661            }
662            "agent_recovery_advice" => {
663                let step = event
664                    .data
665                    .get("step_name")
666                    .and_then(JsonValue::as_str)
667                    .unwrap_or("step");
668                review.push(format!(
669                    "agent recovery advice for {step} (must be validated before re-run)"
670                ));
671            }
672            _ => {}
673        }
674    }
675    let plain = format!(
676        "Safe to automate: {} deterministic event(s) (release analysis, changelog inputs, \
677         successful release steps). Requires human/agent review: {} item(s) including \
678         agent-authored audits, model recovery advice, and any failed deterministic steps. \
679         The crystallized workflow keeps all agent steps behind explicit approval boundaries; \
680         hosts should not promote this candidate to fully autonomous execution without first \
681         resolving the review-required entries.",
682        safe.len(),
683        review.len()
684    );
685    SegmentSummary {
686        deterministic_count: safe.len(),
687        agentic_count: review.len(),
688        safe_to_automate: safe,
689        requires_human_review: review,
690        plain_language: plain,
691    }
692}
693
694/// Build a [`RecoveryFeedbackSummary`] from an ingested fixture. Counts
695/// failed `release_step_recorded` events (the canonical failure record
696/// in `release_harn.harn`) and recovery-advice runs, then emits a
697/// plain-language description of how recovery is represented.
698///
699/// The mirrored `tool_observation` rows for the same release steps are
700/// only consulted for failures whose step name does not already appear
701/// in the deterministic stream — that way a per-step failure is counted
702/// exactly once, but a tool observation that is not paired with a
703/// release-step row (e.g. a one-off shell read failure) still counts.
704pub fn build_recovery_summary(fixture: &ReleaseFixture) -> RecoveryFeedbackSummary {
705    let mut shell_failures = 0usize;
706    let mut failed_steps: Vec<String> = Vec::new();
707    let mut counted_step_names: BTreeSet<String> = BTreeSet::new();
708
709    for event in fixture
710        .deterministic_events
711        .iter()
712        .chain(fixture.tool_observations.iter())
713    {
714        let success = event
715            .data
716            .get("success")
717            .and_then(JsonValue::as_bool)
718            .unwrap_or(true);
719        if success {
720            continue;
721        }
722        let name = event
723            .data
724            .get("name")
725            .or_else(|| event.data.get("step_name"))
726            .and_then(JsonValue::as_str);
727        if let Some(name) = name {
728            if !counted_step_names.insert(name.to_string()) {
729                continue; // already counted from the deterministic stream
730            }
731            if !failed_steps.iter().any(|n| n == name) {
732                failed_steps.push(name.to_string());
733            }
734        }
735        shell_failures += 1;
736    }
737    let recovery_runs = fixture
738        .agent_events
739        .iter()
740        .filter(|event| event.event_type == "agent_recovery_advice")
741        .count();
742    let fed = recovery_runs > 0 && shell_failures > 0;
743    let representation = if fed {
744        format!(
745            "{shell_failures} shell/tool failure(s) were observed and {recovery_runs} of them triggered an `agent_loop` \
746             recovery advice run. Failure stdout/stderr, classification, and recovery hints from \
747             the failing release step were embedded in the model prompt; the model produced \
748             advisory text only — no automatic re-run or live mutation. Reviewers must validate \
749             the advice before re-running the failing step."
750        )
751    } else if shell_failures > 0 {
752        format!(
753            "{shell_failures} shell/tool failure(s) were observed but no agent recovery loop was invoked. \
754             Failure context was recorded for human review only."
755        )
756    } else if recovery_runs > 0 {
757        format!(
758            "No shell/tool failures were observed in this run, but {recovery_runs} agent recovery loop(s) \
759             ran (likely from a previous attempt or a non-shell error path)."
760        )
761    } else {
762        "No shell/tool failures were observed and no recovery loops ran. The release \
763         executed end-to-end on the deterministic path."
764            .to_string()
765    };
766    RecoveryFeedbackSummary {
767        shell_failures_seen: shell_failures,
768        recovery_advice_runs: recovery_runs,
769        failures_fed_into_agent: fed,
770        failed_steps,
771        representation,
772    }
773}
774
775/// Top-level convenience: read a fixture from disk and synthesize a
776/// crystallization candidate ready for bundle-build / validate /
777/// shadow-replay. This is what the CLI subcommand wires up.
778pub fn ingest_release_fixture(
779    dir: &Path,
780    options: CrystallizeOptions,
781) -> Result<
782    (
783        CrystallizationArtifacts,
784        ReleaseFixture,
785        CrystallizationTrace,
786    ),
787    VmError,
788> {
789    let fixture = load_release_fixture(dir)?;
790    let mut trace = release_fixture_to_trace(&fixture);
791    if trace.actions.is_empty() {
792        return Err(VmError::Runtime(format!(
793            "release fixture {} produced zero actions; nothing to crystallize",
794            dir.display()
795        )));
796    }
797    // Stamp a content hash on the trace so the manifest's
798    // source_trace_hashes line up with downstream eval pack rubric
799    // expectations even before normalize_trace re-fills it.
800    let payload = serde_json::to_vec(&trace.actions).unwrap_or_default();
801    trace.source_hash = Some(format!(
802        "sha256:{}",
803        hex::encode(Sha256::digest(payload.as_slice()))
804    ));
805
806    let segment_summary = build_segment_summary(&fixture);
807    let recovery_summary = build_recovery_summary(&fixture);
808    let release = &fixture.manifest.release;
809    let extra_parameters = release_parameter_definitions(release);
810
811    let artifacts = synthesize_candidate_from_trace(
812        trace.clone(),
813        options,
814        extra_parameters,
815        Some(segment_summary),
816        Some(recovery_summary),
817    )?;
818    Ok((artifacts, fixture, trace))
819}
820
821fn release_parameter_definitions(
822    release: &ReleaseFixtureIdentity,
823) -> Vec<WorkflowCandidateParameter> {
824    fn parameter(name: &str, value: &str) -> WorkflowCandidateParameter {
825        WorkflowCandidateParameter {
826            name: name.to_string(),
827            source_paths: vec![format!("manifest.release.{name}")],
828            examples: if value.trim().is_empty() {
829                Vec::new()
830            } else {
831                vec![value.to_string()]
832            },
833            required: true,
834        }
835    }
836    vec![
837        parameter("repo", &release.repo),
838        parameter("base_branch", &release.base_branch),
839        parameter("current_version", &release.current_version),
840        parameter("next_version", &release.next_version),
841        parameter("target_release_branch", &release.target_release_branch),
842    ]
843}
844
845#[cfg(test)]
846mod tests {
847    use super::*;
848    use std::fs;
849    use tempfile::TempDir;
850
851    fn write_fixture(dir: &Path) {
852        let manifest = json!({
853            "schema_version": RELEASE_FIXTURE_SCHEMA,
854            "kind": "release_harn_crystallization_input",
855            "generated_by": "release_harn.harn",
856            "generated_at": "2026-05-02T00:00:00Z",
857            "run_id": "test-run",
858            "source_issue": "https://github.com/burin-labs/harn-bump-fleet/issues/2",
859            "consumer_issue": "https://github.com/burin-labs/harn/issues/1146",
860            "release": {
861                "repo": "/tmp/harn",
862                "mode": "audit",
863                "mock": true,
864                "current_version": "0.7.52",
865                "next_version": "0.7.53",
866                "starting_branch": "main",
867                "target_release_branch": "release/v0.7.53",
868                "base_branch": "main",
869                "latest_tag": "v0.7.52"
870            }
871        });
872        fs::write(
873            dir.join(MANIFEST_FILE),
874            serde_json::to_string_pretty(&manifest).unwrap(),
875        )
876        .unwrap();
877
878        let det = [
879            json!({
880                "schema_version": RELEASE_FIXTURE_SCHEMA,
881                "event_type": "release_analysis",
882                "source": "deterministic",
883                "timestamp": "2026-05-02T00:00:00Z",
884                "data": {
885                    "current_version": "0.7.52",
886                    "next_version": "0.7.53",
887                    "base_branch": "main"
888                }
889            }),
890            json!({
891                "schema_version": RELEASE_FIXTURE_SCHEMA,
892                "event_type": "changelog_audit_inputs",
893                "source": "deterministic",
894                "timestamp": "2026-05-02T00:00:01Z",
895                "data": {"expected_version": "0.7.53"}
896            }),
897            json!({
898                "schema_version": RELEASE_FIXTURE_SCHEMA,
899                "event_type": "release_step_recorded",
900                "source": "deterministic",
901                "timestamp": "2026-05-02T00:00:02Z",
902                "data": {
903                    "name": "ensure-clean-tree",
904                    "command": "git status --porcelain",
905                    "success": true,
906                    "status": 0
907                }
908            }),
909            json!({
910                "schema_version": RELEASE_FIXTURE_SCHEMA,
911                "event_type": "release_step_recorded",
912                "source": "deterministic",
913                "timestamp": "2026-05-02T00:00:03Z",
914                "data": {
915                    "name": "push",
916                    "command": "git push -u origin release/v0.7.53",
917                    "success": false,
918                    "status": 1,
919                    "classification": "branch_protection",
920                    "recovery_hint": "rerun with --no-verify after green hook budget"
921                }
922            }),
923        ];
924        let det_jsonl: String = det
925            .iter()
926            .map(|v| serde_json::to_string(v).unwrap())
927            .collect::<Vec<_>>()
928            .join("\n");
929        fs::write(dir.join(DETERMINISTIC_EVENTS_FILE), det_jsonl + "\n").unwrap();
930
931        let agent = [
932            json!({
933                "schema_version": RELEASE_FIXTURE_SCHEMA,
934                "event_type": "agent_review_attempt",
935                "source": "agent",
936                "timestamp": "2026-05-02T00:00:04Z",
937                "data": {
938                    "index": 0,
939                    "name": "release_audit",
940                    "status": "ok",
941                    "text": "audit passes",
942                    "validation_findings": []
943                }
944            }),
945            json!({
946                "schema_version": RELEASE_FIXTURE_SCHEMA,
947                "event_type": "agent_recovery_advice",
948                "source": "agent",
949                "timestamp": "2026-05-02T00:00:05Z",
950                "data": {
951                    "step_name": "push",
952                    "status": "ok",
953                    "text": "Re-run with --no-verify"
954                }
955            }),
956        ];
957        let agent_jsonl: String = agent
958            .iter()
959            .map(|v| serde_json::to_string(v).unwrap())
960            .collect::<Vec<_>>()
961            .join("\n");
962        fs::write(dir.join(AGENT_EVENTS_FILE), agent_jsonl + "\n").unwrap();
963
964        let tools = [json!({
965            "schema_version": RELEASE_FIXTURE_SCHEMA,
966            "event_type": "tool_observation",
967            "source": "deterministic",
968            "timestamp": "2026-05-02T00:00:06Z",
969            "data": {
970                "tool": "shell",
971                "command": "git status --porcelain",
972                "success": true,
973                "status": 0,
974                "stdout": "",
975                "stderr": ""
976            }
977        })];
978        let tools_jsonl: String = tools
979            .iter()
980            .map(|v| serde_json::to_string(v).unwrap())
981            .collect::<Vec<_>>()
982            .join("\n");
983        fs::write(dir.join(TOOL_OBSERVATIONS_FILE), tools_jsonl + "\n").unwrap();
984    }
985
986    #[test]
987    fn ingest_emits_a_safe_candidate_with_segment_and_recovery_summary() {
988        let temp = TempDir::new().unwrap();
989        write_fixture(temp.path());
990        let (artifacts, fixture, _trace) = ingest_release_fixture(
991            temp.path(),
992            CrystallizeOptions {
993                workflow_name: Some("release_harn".to_string()),
994                package_name: Some("release-harn".to_string()),
995                ..CrystallizeOptions::default()
996            },
997        )
998        .expect("ingest");
999
1000        // Manifest decoded.
1001        assert_eq!(fixture.manifest.release.next_version, "0.7.53");
1002
1003        // Candidate selected and safe to propose.
1004        assert!(artifacts.report.selected_candidate_id.is_some());
1005        let candidate = &artifacts.report.candidates[0];
1006        assert_eq!(candidate.name, "release_harn");
1007        assert!(candidate.shadow.pass);
1008
1009        // Segment summary records both deterministic and agentic items.
1010        let summary = artifacts
1011            .report
1012            .segment_summary
1013            .as_ref()
1014            .expect("segment summary");
1015        assert!(summary.deterministic_count >= 3);
1016        assert!(summary.agentic_count >= 2);
1017        assert!(summary
1018            .requires_human_review
1019            .iter()
1020            .any(|item| item.contains("failed step:push")));
1021
1022        // Recovery summary captures the failed push + advice.
1023        let recovery = artifacts
1024            .report
1025            .recovery_summary
1026            .as_ref()
1027            .expect("recovery summary");
1028        assert!(recovery.shell_failures_seen >= 1);
1029        assert!(recovery.recovery_advice_runs >= 1);
1030        assert!(recovery.failures_fed_into_agent);
1031        assert!(recovery.failed_steps.contains(&"push".to_string()));
1032        assert!(recovery.representation.contains("agent_loop"));
1033
1034        // Release identity surfaces as parameters with `next_version` etc.
1035        let names: Vec<&str> = candidate
1036            .parameters
1037            .iter()
1038            .map(|p| p.name.as_str())
1039            .collect();
1040        assert!(names.contains(&"next_version"));
1041        assert!(names.contains(&"current_version"));
1042        assert!(names.contains(&"base_branch"));
1043
1044        // Generated workflow code references the candidate name.
1045        assert!(artifacts.harn_code.contains("pipeline release_harn"));
1046    }
1047
1048    #[test]
1049    fn rejects_unknown_schema() {
1050        let temp = TempDir::new().unwrap();
1051        let manifest = json!({
1052            "schema_version": "release_harn.crystallization_input.v999",
1053            "kind": "release_harn_crystallization_input",
1054            "generated_by": "release_harn.harn",
1055            "generated_at": "2026-05-02T00:00:00Z",
1056            "run_id": "test-run",
1057            "release": {
1058                "repo": "/tmp/harn",
1059                "mode": "audit",
1060                "mock": true,
1061                "current_version": "0.7.52",
1062                "next_version": "0.7.53",
1063                "starting_branch": "main",
1064                "target_release_branch": "release/v0.7.53",
1065                "base_branch": "main",
1066                "latest_tag": "v0.7.52"
1067            }
1068        });
1069        fs::write(
1070            temp.path().join(MANIFEST_FILE),
1071            serde_json::to_string_pretty(&manifest).unwrap(),
1072        )
1073        .unwrap();
1074        let err = load_release_fixture(temp.path()).unwrap_err();
1075        assert!(format!("{err}").contains("unrecognized schema"));
1076    }
1077
1078    #[test]
1079    fn missing_jsonl_streams_default_to_empty() {
1080        let temp = TempDir::new().unwrap();
1081        let manifest = json!({
1082            "schema_version": RELEASE_FIXTURE_SCHEMA,
1083            "kind": "release_harn_crystallization_input",
1084            "generated_by": "release_harn.harn",
1085            "generated_at": "2026-05-02T00:00:00Z",
1086            "run_id": "no-events-run",
1087            "release": {
1088                "repo": "/tmp/harn",
1089                "mode": "audit",
1090                "mock": true,
1091                "current_version": "0.7.52",
1092                "next_version": "0.7.53",
1093                "starting_branch": "main",
1094                "target_release_branch": "release/v0.7.53",
1095                "base_branch": "main",
1096                "latest_tag": "v0.7.52"
1097            }
1098        });
1099        fs::write(
1100            temp.path().join(MANIFEST_FILE),
1101            serde_json::to_string_pretty(&manifest).unwrap(),
1102        )
1103        .unwrap();
1104        let fixture = load_release_fixture(temp.path()).expect("load");
1105        assert!(fixture.deterministic_events.is_empty());
1106        assert!(fixture.agent_events.is_empty());
1107        assert!(fixture.tool_observations.is_empty());
1108
1109        // Without any events, ingest fails closed (no actions to crystallize).
1110        let err = ingest_release_fixture(temp.path(), CrystallizeOptions::default()).unwrap_err();
1111        assert!(format!("{err}").contains("zero actions"));
1112    }
1113}