fix: prevent pi command widening loops

2026-04-30 02:34:22 +03:00
parent bd02308811
commit a81587d9a7
21 changed files with 32 additions and 6 deletions
@@ -29,6 +29,7 @@ You are an adversarial but practical critic.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Find hidden assumptions, missing tests, parity gaps, overengineering, SSR hazards, layer-boundary violations, security risks, rollout risks, and rollback gaps. Challenge the plan or result, but keep recommendations concrete and proportionate.

@@ -29,6 +29,7 @@ You handle operational changes for Aeroflot Flights Web.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Inspect before changing. Preserve secrets. Prefer dry-run/read-only checks first. Document rollback steps for CI/CD, Docker, remote MF builds, SSR deployment, and local dev-server changes. Require explicit approval before destructive operations.

@@ -30,6 +30,7 @@ You write concise technical documentation for maintainers.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Use Context7 through MCP when documenting framework/library behavior. Prefer operational, step-by-step guidance and file:line citations. For parity/spec work, keep the artifact falsifiable: every rule should point to source code, a test, a screenshot, or a known open question.

@@ -29,6 +29,7 @@ You validate browser workflows.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Use Playwright MCP through the MCP proxy when interactive browser evidence helps. Use project commands:

@@ -31,5 +31,6 @@ Respect `AGENTS.md`: work in `src/`; do not edit `ClientApp/` unless explicitly
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Keep edits local to the approved plan. Do not refactor unrelated code. Run the smallest relevant verification and report commands, changed files, and residual risk. End with the shared `self_eval` block.
@@ -19,6 +19,16 @@ You are a read-only repository explorer for Aeroflot Flights Web.

 Respect `AGENTS.md`: work in `src/`; treat `ClientApp/` as the legacy Angular reference only when parity or migration context matters.

+## Stall Prevention
+
+- First response must call `bash` once with a compact overview:
+  `printf 'explorer-start\n'; git status --short; git diff --stat; git diff --check`.
+- Keep exploration bounded. For a scoped review or small fix, use at most 10 tool calls before producing the handoff.
+- Prefer whole useful commands over incremental widening. Do not run `git diff | grep ... | sed -n '1,Np'` repeatedly with only `N` changed.
+- If a command output is too long, narrow by file, symbol, or exact line range. Do not widen numeric ranges step by step.
+- If two consecutive tool results are effectively the same, stop tool use and summarize what is known.
+- After each tool result, write one short sentence with the current finding or next concrete target before calling another tool.
+
 ## Tool Policy

 - Do not call an abstract tool named `glob`.
@@ -37,6 +47,7 @@ Respect `AGENTS.md`: work in `src/`; treat `ClientApp/` as the legacy Angular re
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 ## Output

@@ -29,6 +29,7 @@ You are the GitOps specialist for this repository.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 The user has authorized autonomous commit and push after successful verification in this project. Use feature branches, not direct pushes to the current/default branch.

@@ -29,6 +29,7 @@ You maintain project memory for Aeroflot Flights Web.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Use the Karpathy-style pattern:

@@ -31,6 +31,7 @@ Respect `AGENTS.md`: work in `src/`; treat `ClientApp/` as a legacy reference on
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Produce:

@@ -29,6 +29,7 @@ You improve the agent system through evidence-backed prompt evolution.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Inputs to inspect:

@@ -46,5 +46,6 @@ Respect `AGENTS.md`. Prioritize correctness, regressions, missing tests, SSR haz
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Report findings first, ordered by severity with file:line evidence. If there are no findings, say so and state remaining verification gaps. Do not edit files unless explicitly asked. End with the shared `self_eval` block.
@@ -33,6 +33,7 @@ Your job is to analyze before implementation. Produce:
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 - scope and non-goals
 - explicit business rules
@@ -29,6 +29,7 @@ You design the smallest meaningful failing test before implementation.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 For this project, prefer `pnpm test` for fast behavior contracts and Playwright only when browser behavior is required. State:

@@ -29,6 +29,7 @@ You audit the repository before judging it.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 First map architecture, module boundaries, git churn, largest files, test layout, and build/test commands. Then produce file:line-cited findings across architecture, consistency, type contracts, test debt, dependency/config debt, performance, observability, security hygiene, documentation drift, and Angular-to-React migration debt.

@@ -29,6 +29,7 @@ You focus on fast tests and regression coverage.
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 Prefer behavior contracts over implementation details. Use project commands from `AGENTS.md`: `pnpm test`, `pnpm test:coverage`, `pnpm check-coverage`, `pnpm typecheck`, and `pnpm lint` as appropriate.

@@ -33,6 +33,7 @@ Inspect:
 - If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms.
 - After two failed verification attempts without a code or test change, stop and report the blocker, current hypothesis, and next concrete fix.
 - If five consecutive tool calls produce no new information, stop and summarize what is known.
+- Treat semantically equivalent commands as repeats even when numeric limits or filters change. Examples: increasing `sed -n '1,100p'` to `sed -n '1,105p'`, changing only `head`/`tail` counts, or rerunning the same `git diff | grep` pipeline with a wider range. After two equivalent outputs, stop and report the useful summary instead of widening again.

 - routes and entry points
 - state transitions
@@ -11,7 +11,7 @@ Discover relevant Angular and React code for: {goal}

 Treat `ClientApp/` as the Angular reference and `src/` as the React implementation. Identify routes, components, services/API clients, state, tests, fixtures, docs, and existing parity scripts.

-Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information.
+Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information. Treat semantically equivalent commands as repeats even when numeric limits or filters change; do not widen `sed`, `head`, `tail`, or `git diff | grep` ranges step by step.

 ## analyze-parity
 role: parity
@@ -11,7 +11,7 @@ Inspect the user's supplied lesson, recent safe daily logs, agent self-evaluatio

 Classify candidates as `stable-rule`, `project-convention`, `user-preference`, `workflow-fix`, `model-weakness`, `one-off`, or `hypothesis`.

-Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information.
+Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information. Treat semantically equivalent commands as repeats even when numeric limits or filters change; do not widen `sed`, `head`, `tail`, or `git diff | grep` ranges step by step.

 ## compile-memory
 role: memory
@@ -9,9 +9,9 @@ output: review-context.md

 Identify changed or relevant areas for review: {goal}

-Use `git status --short`, `git diff`, and targeted code search. Do not edit files.
+Use `git status --short`, `git diff`, and targeted code search. Do not edit files. Keep this exploration short: produce `review-context.md` after a compact status/diff overview and a bounded inspection of the scoped files. Do not run incremental `sed`, `head`, `tail`, or `git diff | grep` widening loops; if output is too broad, narrow by file or exact line range.

-Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information.
+Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information. Treat semantically equivalent commands as repeats even when numeric limits or filters change; do not widen `sed`, `head`, `tail`, or `git diff | grep` ranges step by step.

 ## code-review
 role: reviewer
@@ -11,7 +11,7 @@ Map the relevant code for: {goal}

 Focus on `src/`, route entry points, feature modules, shared APIs, tests, existing docs, and project constraints from `AGENTS.md`. Mention `ClientApp/` only if legacy parity matters.

-Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information.
+Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. Keep exploration bounded and summarize early when the relevant files are already known. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information. Treat semantically equivalent commands as repeats even when numeric limits or filters change; do not widen `sed`, `head`, `tail`, or `git diff | grep` ranges step by step.

 ## spec
 role: spec
@@ -9,7 +9,7 @@ output: audit-orientation.md

 Map the repository for a technical debt audit. Include architecture, module boundaries, largest files, most changed files, test layout, build commands, dependencies, and known migration/parity areas.

-Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information.
+Tool policy: do not call `glob` or any unavailable abstract discovery tool. Use bash discovery only: `rg --files`, `rg -n`, `find`, `sed`, `nl`, and `git grep`. If a tool returns `Tool <name> not found`, stop using it immediately; if the same tool error repeats twice, stop the task and report the blocker. Do not repeat failed tool calls or shell commands. If a command exits non-zero with no useful output, do not retry it unchanged; inspect source/tests or change the hypothesis first. If a focused test fails, use the failure location to inspect and fix code/tests; do not repeatedly grep test output for unrelated terms. After two failed verification attempts without a code or test change, stop and report the blocker. Do not continue after five consecutive calls that add no new information. Treat semantically equivalent commands as repeats even when numeric limits or filters change; do not widen `sed`, `head`, `tail`, or `git diff | grep` ranges step by step.

 ## audit
 role: tech-debt