Feat/step76 89 sega pipeline

.gitignore

@@ -178,6 +178,18 @@ metagpt/roles/idea_agent.py
 
 # output folder
 output
+# Step 76+ results live in output2 (user request: keep new runs separate from
+# the crowded output/); the bare `output` rule above does not match it.
+# Ignore the RESULTS but track the drivers -- otherwise the code that produced
+# a paper number is unreviewable. Use `output2/*` (not `output2/`) so the
+# negation works: git will not descend into an ignored directory.
+layout_agent/output2/*
+!layout_agent/output2/*.py
+# RUN_METADATA.json is the reproducibility freeze record for a run: which
+# commit, which model, which sample ids. Tiny, and worthless if not tracked.
+!layout_agent/output2/*/
+layout_agent/output2/*/*
+!layout_agent/output2/*/RUN_METADATA.json
 tmp.png
 .dependencies.json
 tests/metagpt/utils/file_repo_git

AGENTS.md

@@ -0,0 +1,120 @@
+<INSTRUCTIONS>
+# User-visible execution protocol
+
+These rules are persistent project instructions and apply to every future session.
+
+## Explain before acting
+
+- Before starting tools or a long-running command, tell the user in plain language:
+  what is being run, why it is needed, whether it uses paid APIs, the expected
+  duration, and the exact condition that will stop or unblock the next step.
+- Keep the user oriented. Every status update must say what is happening now and
+  what will happen next; do not expose only internal agent/reviewer mechanics.
+- For a command expected to take more than two minutes, provide one start
+  checkpoint and then report only meaningful milestones, completion, failure, or
+  when the user asks. Do not spend Codex usage on repetitive heartbeat polling.
+
+## Conserve Codex usage
+
+- Use the minimum number of agents and review rounds needed for the risk. Do not
+  launch repeated implementation/review loops after the acceptance criteria pass.
+- Prefer one implementation pass plus one focused verification pass. Add another
+  reviewer only for a concrete unresolved risk, not as a routine ritual.
+- If the user asks to reduce usage, immediately stop nonessential agents, polling,
+  and commentary. Never interrupt a useful local process merely to inspect it.
+
+## Final-only handoff and Git persistence
+
+- Do not update `layout_agent/next_step.md` after individual commands, checks,
+  milestones, or intermediate discoveries. Keep intermediate state in the active
+  session instead of creating repeated handoff churn.
+- Update `layout_agent/next_step.md` exactly once, immediately before handing
+  control back after execution completes, genuinely blocks, or reaches a paid
+  authorization boundary. That consolidated update must include the timestamp,
+  material commands and results, artifacts, paid/API cost, remaining work, and
+  safest resume command.
+- After the consolidated handoff, run one proportionate final verification pass,
+  create one scoped commit containing only the task files and final handoff, and
+  push once. Do not create a second handoff-only commit merely to record the first
+  commit hash or push receipt; report those in the final response instead.
+- Read-only answers that do not materially advance execution require no handoff,
+  commit, or push.
+- Never include unrelated pre-existing dirty files. Never stash, reset, clean, or
+  overwrite user work to make a commit possible.
+- Always report the branch, commit hash, push result, checks run, and any files
+  intentionally left uncommitted. If commit or push is blocked, state the exact
+  blocker instead of silently leaving changes local.
+
+## Cost and completion guardrails
+
+- Never run a paid API, LLM judge, or paid generation step without first stating
+  the exact call/token budget and receiving explicit paid-run authorization.
+- A completion report must clearly list: what changed, what ran, the result and
+  artifact, remaining work, `next_step.md` status, and commit/push status.
+</INSTRUCTIONS>
+
+<claude-mem-context>
+# Memory Context
+
+# [MetaGPT] recent context, 2026-07-12 3:14am GMT+8
+
+Legend: 🎯session 🔴bugfix 🟣feature 🔄refactor ✅change 🔵discovery ⚖️decision
+Format: ID TIME TYPE TITLE
+Fetch details: get_observations([IDs]) | Search: mem-search skill
+
+Stats: 50 obs (20,187t read) | 1,752,582t work | 99% savings
+
+### Jul 12, 2026
+4988 2:26a 🟣 evaluate_a3_sega.py — Atomic RENAME_NOREPLACE Publication and Full Bundle Validation Before Write
+4989 " 🟣 validate_evaluation_bundle() — Full v1 Sidecar Contract Enforcement in a3_sega_evaluator.py
+4990 " 🟣 Matched Sample ID Enforcement and Code/Detector Rehash in evaluate_a3_sega.py
+4996 2:28a 🔵 phase1_hardening Callback Pattern — Third Brief Signal at 18:26:25, Then Silent Again
+4997 2:30a 🔴 Orchestrator Pre-Stages Re-Verification Round While phase1_hardening Still Running
+4998 2:31a 🔄 saliency_basnet_isnet.py — Full Hardening: Snapshot Resolution, Identity Recording, Output Validation, Strict ISNet Identity
+4999 " 🟣 evaluate_a3_sega.py — Pre-Publish Lineage Check Callback Added to _write_results
+5000 " 🟣 test_sega_metrics.py Expanded — PKU Quirk Tests and Overlay Denominator Tests Added
+5001 " ✅ Phase 1 Hardening Final State — 64 Tests Passing, All Checks Clean, Smoke-2 Confirmed at 02:28 CST
+5002 2:32a 🔵 phase1_hardening Entered Another Silent Period — Completion Estimate Was Premature
+5003 2:33a 🔵 new_plam.md A3 Task Audit — Phase 3 Remaining Items After Phase 1+2 Completion
+5004 2:34a 🔵 Pipeline Enters Final Re-Verification Phase — Three Rapid Consecutive Completions at 18:33–18:34
+5005 2:36a 🟣 A3 SEGA Evaluator Phase 1 Hardening — Major Expansion to 1,157 Lines
+5006 " 🟣 A3 SEGA Test Suite Expanded — 64 Tests Passing (Up from 42)
+5007 " 🔵 BASNet + ISNet Detector Artifact Verification — Hashes and Session Identity Confirmed
+5008 " ✅ next_step.md Updated — Formal N=100 Run Blocked Pending Independent Review
+5009 2:37a 🔵 Orchestrator Retrying Identical followup_task to phase1_hardening — Subtask Not Yet Acknowledged
+5010 " 🔵 new_plam.md A3 Task Audit — Phase 3 Remaining Items After Phase 1+2 Completion
+5011 2:39a 🔵 phase1_hardening Sixth Callback at 18:38:28 — Still Active at 23-Minute Mark
+5012 2:40a 🔵 phase1_hardening Silent Period Continues Past 25-Minute Total Runtime
+5014 2:43a 🔵 phase1_hardening Seventh Callback at 18:41:55 — Orchestrator on ~100s Retry Loop
+5015 " 🔵 A3 SEGA Phase 1 Pipeline Wall-Clock Cost Profile — 28+ Minutes for Hardening Alone
+5016 2:44a 🔵 phase1_hardening Past 29-Minute Mark — Longest Agent Execution in A3 SEGA Pipeline
+5018 2:46a 🔵 phase1_hardening Eighth Callback at 18:45:22 — Orchestrator Cycling Through Multiple Nudge Messages
+5019 2:48a 🔵 Orchestrator Sent Fourth Distinct Escalation Payload to phase1_hardening at 18:48:33
+5020 2:52a 🔵 phase1_hardening Ninth Callback at 18:48:47 — 33+ Minutes Runtime, Still No Completion
+5021 2:54a 🔵 Orchestrator Shortened Timeout Windows to 30s and Sent Fifth Escalation Payload at 38-Minute Mark
+5023 " 🔵 phase1_hardening Silent for 6+ Minutes — Longest Quiet Period in Session
+5024 2:55a 🔵 phase1_antipattern Final Report — No Blockers, 3 Non-Blocking Edge Cases, renameat2 and ISNet Fixes Confirmed
+5025 " 🔵 phase1_final_quality Report — 4 P1 Blockers Prevent Phase 2 Unlock; 64/64 Tests Pass
+5030 " 🔵 phase1_hardening Force-Interrupted After 40 Minutes — Orchestrator Issued interrupt_agent
+5031 2:58a 🔵 Post-Interrupt Pivot — Orchestrator Re-Engaged All Three Audit Agents with New P1 Blocker Tasks
+5032 " 🔵 new_plam.md A3 Task Audit — Phase 3 Has Three Remaining Items; Phase1_final_quality Reports 4 P1 Blockers
+5033 3:00a 🔵 Post-Interrupt Re-Verification Phase Active — Split Wait Results Confirm Multiple Agents Running
+5034 " 🔵 Phase 1 Final Gate Verification — 84 Tests Pass, Real Runtime Contract Valid, ISNet Artifact Hashes Confirmed On-Disk
+5035 " 🔵 Post-Interrupt Audit Round In Progress — Awaiting list_agents for P1 Blocker Resolution Results
+5036 " 🔵 Frozen Detector Artifact Hashes Confirmed — ISNet 178MB ONNX + BASNet Revision Pinned and Schema-Valid
+5037 3:02a ⚖️ Phase 1 Declared Complete — Full 5-Phase A3 SEGA Evaluation Roadmap Established
+5038 " 🟣 phase1_commit and phase1_sync Agents Spawned — Phase 1 Work Being Persisted to Git
+5039 3:04a 🟣 Phase 2 Formal Evaluation Started — phase2_run Agent Spawned for T0/T2/T3 Zero-LLM Six-Axis Run
+5040 " 🔵 new_plam.md A3 Task Audit — Phase 3 Has Three Remaining Items
+5041 3:06a 🔵 phase2_run Agent Active and Responding — Formal Evaluation Execution Confirmed Started
+5043 3:07a 🟣 A3 SEGA Phase 2 Real-Detector Smoke — Passed with Metric Values
+5044 " 🟣 Formal A3 SEGA Relation N=100 Evaluation Launched — evaluation-id a3-relation-n100-t0-t2-t3-sega-v1
+5045 " 🔵 BASNet Safetensors Load — Non-Meta Parameter Warnings Are Cosmetic Only
+5042 " 🔵 phase2_run In Silent Execution — Formal Evaluation Processing T0/T2/T3 Samples
+5046 3:09a 🔵 phase2_run Callback Cadence Established — ~2-Minute Inter-Callback Intervals During Evaluation
+5047 3:10a 🔵 Formal A3 SEGA N=100 Evaluation — Runtime Resource Profile and Write-at-End Architecture Confirmed
+5048 3:11a 🔵 New wait_agent Outcome Observed — "Wait interrupted by new input" Indicates Concurrent Agent Activity
+5049 3:12a 🔵 A3 SEGA Evaluator Process Profile — 56 Threads, Per-Sample File Open/Close, Stable Memory Growth
+
+Access 1753k tokens of past work via get_observations([IDs]) or mem-search skill.
+</claude-mem-context>