WIP: Harness resilience — relaxed phase-2 gate, enriched resume diagnostics, timestamped summaries

.opencode/agents/auditor.md

@@ -40,10 +40,6 @@ Use target-specific skills when they apply, for example:
 - `.opencode/skills/c-cpp-security/SKILL.md`
 - `.opencode/skills/juliet-benchmark/SKILL.md`
 
-Also reference when writing run summaries:
-
-- `templates/run-summary.md`
-
 ## Mission
 
 Create concrete vulnerability hypotheses.
@@ -240,11 +236,11 @@ If the target contains C/C++:
 
 Before finishing:
 
+- The phase prompt specifies required durable artifacts. Follow it precisely.
 - each finding is in `itemdb/findings/PENDING/`,
 - each finding uses valid frontmatter,
 - each finding has a unique id,
 - each finding has a specific validation plan,
 - no finding is marked confirmed,
 - no vague findings were created,
-- duplicates were considered,
-- a short run summary is written when practical.
+- duplicates were considered.

.opencode/agents/exploiter.md

@@ -28,7 +28,6 @@ Before developing an exploit, read:
 - the assigned CONFIRMED finding
 - existing validation evidence under `itemdb/evidence/<finding-id>/`
 - relevant source files under `src/`
-- `templates/run-summary.md`
 - sandbox documentation under `sandbox/`
 
 Use target-specific skills when they apply, for example:
@@ -252,6 +251,7 @@ Bad:
 
 Before finishing:
 
+- The phase prompt specifies required durable artifacts. Follow it precisely.
 - exploitation was attempted for the assigned finding,
 - `itemdb/evidence/<finding-id>/exploits/` directory exists,
 - `exploits/README.md` exists with exploitation details,
@@ -267,5 +267,4 @@ Before finishing:
   `# Inputs and preconditions`, `# Recording`, and `# Remediation idea`
   (with corrected-code excerpt or diff) are filled,
 - finding is in the correct status directory (EXPLOITED or CONFIRMED),
-- limitations are documented,
-- a run summary is written when practical.
+- limitations are documented.

.opencode/agents/recon.md

@@ -22,10 +22,6 @@ Before starting reconnaissance, read:
 - `.opencode/skills/source-recon/SKILL.md`
 - `.opencode/skills/sandbox-bootstrap/SKILL.md`
 
-Also reference when writing run summaries:
-
-- `templates/run-summary.md`
-
 Do not load target-specific security skills before first mapping the source
 tree broadly. After broad structural mapping, you may consult target-specific
 skills only for reconnaissance checklists, not for vulnerability deep dives or
@@ -342,15 +338,15 @@ Before finishing:
 
 Phase 1a:
 
+- The phase prompt specifies required durable artifacts. Follow it precisely.
 - all required notes exist,
 - target type is stated with confidence,
 - attack surfaces are listed,
 - build and execution model are documented,
 - validation model is documented,
 - interesting files are listed,
 - uncertainty is documented,
-- no low-quality findings were created,
-- a run summary is written to `runs/` using `templates/run-summary.md`.
+- no low-quality findings were created.
 
 Phase 1b:
 

.opencode/agents/reviewer.md

@@ -26,10 +26,6 @@ Before reviewing findings, read:
 - all findings under `itemdb/findings/PENDING/`
 - related findings under `itemdb/findings/CONFIRMED/`, `REJECTED/`, and `DUPLICATE/`
 
-Also reference when writing run summaries:
-
-- `templates/run-summary.md`
-
 Use target-specific skills when they apply, for example:
 
 - `.opencode/skills/c-cpp-security/SKILL.md`
@@ -209,11 +205,11 @@ For C/C++ findings:
 
 Before finishing:
 
+- The phase prompt specifies required durable artifacts. Follow it precisely.
 - every reviewed finding has an updated `# Counter-analysis`,
 - each finding is in the correct status directory,
 - confidence is adjusted when appropriate,
 - validation plans are improved where needed,
 - rejected findings have clear rejection reasons,
 - duplicate findings reference canonical ids,
-- no finding is marked CONFIRMED (confirmation belongs to Phase 4 validator),
-- a run summary is written when practical.
+- no finding is marked CONFIRMED (confirmation belongs to Phase 4 validator).

.opencode/agents/validator.md

@@ -28,7 +28,6 @@ Before validating a finding, read:
 - the assigned finding
 - relevant source files under `src/`
 - `templates/evidence-readme.md`
-- `templates/run-summary.md`
 - sandbox documentation under `sandbox/`
 
 Use target-specific skills when they apply, for example:
@@ -387,12 +386,12 @@ For benchmark targets:
 
 Before finishing:
 
+- The phase prompt specifies required durable artifacts. Follow it precisely.
 - evidence directory exists if validation was attempted,
 - commands are recorded,
 - observed results are recorded,
 - finding frontmatter is updated,
 - validation result section is updated,
 - evidence section references files,
 - finding is moved to the correct status directory,
-- unresolved limitations are documented,
-- a run summary is written when practical.
+- unresolved limitations are documented.