fix(context): dedupe subdirectory hints by content digest and skip backup/vendor dirs

[BKStock]

Problem

SubdirectoryHintTracker dedupes by directory, never by content. The same AGENTS.md reachable through more than one path is injected once per path.

This is not hypothetical — three common setups alias a single file:

• symlinked shared workspaces (one agent workspace whose context files are symlinks into another tree)
• hardlinks
• timestamped backup copies (backups/2026-07-27/AGENTS.md)

There is also no exclusion list, so directories that only ever hold copies — backups/, node_modules/, venv/, site-packages/, .git/, .Trash/, vendor/ — are scanned like real project directories. Taking a backup of a context file permanently inflates every subsequent session.

Measured on a session that touched a symlinked workspace plus a backup directory: 3 injections, ~24,000 chars, all three byte-identical.

Change

Content digest. Every injected hint body is recorded as a sha256. Repeat content is skipped regardless of the path it arrived through. The working directory's own context file is seeded at construction, so the copy prompt_builder already loaded at startup is recognised as a duplicate rather than sent a second time.

Excluded directories. A frozenset of directory names that hold copies rather than authoritative context. Screening is relative to working_dir, so a project that legitimately lives under vendor/ keeps discovering its own subdirectory hints — only segments below the working dir are checked.

Same session after the change: 1 injection, 8,112 chars.

Tests

14 added (18 → 32), covering:

• symlink aliasing, byte-identical copies, working-dir seeding
• genuinely different content still injected (dedupe must not over-suppress)
• each excluded directory name, parametrized
• excluded ancestors block descendants
• working dir inside an excluded name still discovers its own subdirs

155 passed, 1 skipped across test_subdirectory_hints.py + test_prompt_builder.py.

Run directly with python -m pytest rather than scripts/run_tests.sh — the wrapper's venv probe is broken on this checkout, fixed separately in #72447.

[teknium1]

Thanks for targeting a real context-bloat path: current main only remembers visited directories in agent/subdirectory_hints.py:72-74, while both executor paths append tracker output to tool results (agent/tool_executor.py:1210-1217, 1884-1889).

Problems

• The new working-directory seed is unconditional, but project context is only built when skip_context_files is false (agent/system_prompt.py:482-498). It can therefore suppress an identical subdirectory hint that was never injected. It also does not match startup precedence: .hermes.md is selected before AGENTS.md (agent/prompt_builder.py:2169-2175).
• The vendor and third_party exclusions suppress all descendant hints below those names. That changes the documented hierarchical discovery behavior for an owned nested component; the PR only tests the separate case where the working directory itself is inside vendor.

Suggested changes

• Seed only the actual startup-selected context source, and only when context-file injection is enabled; cover skip_context_files and .hermes.md precedence.
• Narrow or qualify the broad directory-name exclusions and cover an owned project nested beneath vendor.

Automated hermes-sweeper review.

[teknium1]

This seed is unconditional, but agent/system_prompt.py:482-498 skips project-context injection when skip_context_files=True; in that mode an identical subdirectory hint has not already been delivered and must not be suppressed. It also needs to follow the .hermes.md-first startup precedence in agent/prompt_builder.py:2169-2175.

[teknium1]

vendor is not reliably a copy-only directory: this suppresses project/vendor/component/AGENTS.md even when the agent is working in that owned nested component. The current exemption only covers a working directory already inside vendor; please preserve or explicitly scope discovery for nested projects.

[kshitijk4poor]

Thanks @BKStock — the gap you fixed is not only still open on current main, we CONFIRMED THE LIVE SYMPTOM in our own triage sessions today (the same .worktrees/*/AGENTS.md injected multiple times per session). Salvaged with your authorship in #77787: both groups ported (vendor-dir skip + sha256 digest dedupe), your dead outside-workdir branch dropped since main now rejects those paths earlier (f4953bc), tests merged into main's existing file, both mechanisms mutation-checked. Closing in favor of the salvage.