fix: edit-loop guard + self-authored artifact tagging

[manolitnora]

Why

Reasoning-pattern guards extracted from a transcript where the model spent ~30 turns optimizing a fabricated analytical framework against a self-confirming test it had written itself.

Two failure modes addressed

1. Edit loops

Model edited the same file 8–10 times in succession, tweaking threshold values back and forth. Each iteration emitted a 'Critical Finding' summary. No fundamental change between iterations — pure churn.

ToolExecutionContext now carries an edit_history: dict[str, int]. After 5 writes/edits to the same resolved path within one context lifetime, _track_write_and_check_loop refuses with ToolExecutionError:

'edit-loop guard: refused to write/edit X a 6-th time in this session (limit 5). This pattern indicates a tweak-and-rerun loop, not progress. Stop and explain to the user what fundamental change you expect to be different — or what hypothesis you are testing — before editing this file again.'

Permanent for the context's lifetime. Cannot be talked past.

2. Self-authored measurements

Model wrote retrieval_accuracy_test.py whose queries were sampled from the very pattern files it was testing against, then cited '0.6% combined accuracy' as evidence of a critical finding. That number was test-design noise. Self-grading.

ToolExecutionContext now carries self_authored_paths: set[str]. When _read_file reads a path the agent wrote earlier in the session, prepend:

'[self-authored: this file was written or edited by the agent earlier in this session. Results from it are not independent measurements. Treat with skepticism.]'

Applies to full-file reads and line-range reads.

Tests

6 new tests in tests/test_edit_loop_and_self_authored.py:

• Repeated writes to same file eventually refused
• Writes to different files don't share a counter
• Mixing _write_file and _edit_file counts together
• Read after self-write includes warning header
• Read of external file has no warning
• Warning preserved with line-range reads

Full suite: 1488 passed.

Limits

• Does NOT prevent the model from writing a self-confirming test in the first place — only flags it on subsequent reads.
• Does NOT catch markdown-churn (each summary is a different path, so the per-path counter doesn't trip).
• Does NOT close the reasoning-layer gap entirely. Model can still ignore the warning header and proceed as if the result were objective. The header is information, not enforcement.

Test plan

• Reviewer pulls and runs the new test file.
• Reviewer confirms _EDIT_LOOP_LIMIT = 5 is the right threshold for their workflows. Five gives room for normal iterative work; raising to 8–10 may be appropriate for some teams.

🤖 Generated with Claude Code

[manolitnora]

Updated this branch with two more guards (commit 5358abd):

Bash self-authored banner — when the agent runs a script it wrote earlier in the session (e.g., python3 retrieval_test.py), the captured output is prefixed with [self-authored: ... not an independent measurement. Treat with skepticism.]. Mirrors the read-time warning for execution-time results, since the prior commit only caught _read_file of self-authored paths.

Markdown-churn guard — refuses further summary-shape .md writes after 4 in one context. Catches the pattern where each iteration emits CRITICAL_FINDING_*.md, SESSION_SUMMARY_*.md, QUICK_REFERENCE.md, gaps_filled_*.md etc. as fake progress markers. Each is a different path so the per-path edit-loop limit doesn't trip — cumulative count does. Allowlists README / CHANGELOG / LICENSE / CONTRIBUTING and the docs/ directory.

9 new tests, full suite 1497 passed.

[manolitnora]

Heads-up: I noticed .latti/ (104 local agent runtime files) had also been accidentally committed into this branch — same root cause as PR #32. Just force-pushed an untrack .latti/ + gitignore commit to clean it. The actual fix in this PR is unchanged; the diff should now look much smaller. Sorry for the noise.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.