… tighten dispatch

Sparse-history dispatches to ce-session-historian were running 17+ minutes
with 33 tool calls before stream-idle-timing-out, when the correct answer
("no relevant prior sessions") should arrive in under a minute. Six gaps
were compounding:

- No skill primitive for "search inventory by keyword across sessions",
  forcing the agent to roll 20 by-hand `grep -l` invocations across MB-sized
  JSONL files.
- Soft "typically 2-5 sessions" guidance with no hard cap.
- Tail-after-head extraction allowed unconditionally; ran reflexively on
  half of selected sessions.
- Verbose dispatch prompt from /ce-compound with topic-keyword bullets that
  licensed the agent to keep widening.
- Repo name not pre-resolved by caller; agent burned its first turn deriving
  via git rev-parse.
- No wall-time budget anywhere in the agent.

This change:

- Adds an opt-in --keyword K1[,K2,...] mode to ce-session-inventory's
  extract-metadata.py. When set, the script does a full-file case-insensitive
  substring scan, filters out zero-match sessions, and emits per-session
  match_count plus per-keyword counts. _meta gains files_matched.
- Tightens ce-session-historian: 5-7 min wall budget, hard cap of 5
  deep-dives, conditional tail-extract, gitBranch start-of-session
  limitation note, Step 3 #4 now points to --keyword instead of by-hand grep.
- Tightens /ce-compound: pre-resolves repo + branch via backtick syntax,
  rewrites the historian dispatch block as a 5-field schema (pre-resolved
  context, time window, problem topic, filter rule, output schema).

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

…d evals scaffolding

The first iteration on session-historian (this PR's prior commit) tightened
prose around extraction caps, conditional tail extraction, and a keyword-search
hint, but did not stop the agent's strongest default behavior: when the
candidate list is small enough to fit under the cap, the agent extracts every
returned session "to verify" rather than running the keyword filter first. On
sparse-mismatch dispatches that means deep-extracting 3 unrelated sessions
instead of issuing one --keyword invocation that returns 0 matches in a single
call.

This commit replaces the soft Step 3 priority list with an explicit numbered
decision sequence:
- Branch filter first.
- If branch filter is empty, run ce-session-inventory --keyword with keywords
  derived from the dispatch problem topic.
- If files_matched is 0, return "no relevant prior sessions" and STOP --
  ce-session-extract is not invoked.

Step 4 gains an explicit "only run if Step 3 produced selected sessions"
guard, and a new guardrail at the top forbids extraction-to-verify outright.
The time-budget block drops the 5-7 minute target wording (which read as a
target rather than a max) in favor of "stop when complete; structural caps
bound runtime by construction."

Adds top-level evals/ for repo-only LLM-driven behavioral checks. The
session-historian eval covers the sparse-history scenario via synthetic
~/.claude/projects/-tmp-eval-... fixtures and a generic-subagent dispatch
pattern (inject the agent definition into a general-purpose subagent rather
than dispatching the typed agent, which reads cached definitions from
session start). Documents the cache caveat in evals/README.md.

Validated on the sparse-mismatch scenario:
- Before: 4 tool calls (inventory + 3 deep extracts), agent ignored keyword
  filter even with explicit prose guidance.
- After: 2 tool calls (inventory + --keyword filter), zero deep extractions,
  correct response. Wall time 55s, well under the 60s soft target. Full
  results in evals/session-historian/results-2026-04-25.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

…scaffolding

Two changes that emerged from validating the prior commits:

1. --keyword was scanning JSONL metadata, not user/assistant content. Common
   topic words like "session" matched every file via the sessionId field —
   on a 4-session keyword-match fixture, only 1 had real topical content,
   but all 4 returned non-zero match_count from metadata noise. extract-metadata.py
   now extracts user message text + assistant text blocks first and scans
   only those, skipping JSONL field names, tool_use blocks (tool names + inputs),
   tool_result blocks, and thinking blocks. Verified on the same fixture:
   files_matched drops from 4 to 1, only the real match remains.

   Adds 3 tests:
   - sessionId / gitBranch / parentUuid as keywords return zero matches
     against the Claude fixture (would have all matched under the prior impl)
   - "Edit" as a keyword does not match against tool_use names in the fixture
   - "auth" still matches against actual user/assistant text content

2. Drops the evals/ scaffolding added in earlier commits. The skill-creator
   skill is the canonical tool for evaluating agent and skill changes — it
   has its own conventions (evals/evals.json, <skill>-workspace/) and a
   purpose-built workflow. The repo-local scaffolding under evals/ was a
   one-off investigation artifact that doesn't conform to skill-creator's
   shape and would silently rot when agent definitions change.

   The durable lessons from that work are kept in repo-root AGENTS.md
   (skill-creator as the canonical path; agents AND skills both cache at
   session start; never edit ~/.claude/plugins/ to test). Removes the
   evals/ entry from the AGENTS.md Directory Layout section in the same
   commit so the directory reference doesn't outlive the directory.

The two positive-path scenarios (branch-match success, keyword-match without
branch-match) were both validated during this work via the skill-creator
dispatch pattern. Results captured in the PR description and commit history;
they don't need a committed results doc.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

- Pre-resolved repo name: distinguish absolute vs relative output from
  git rev-parse --git-common-dir. The previous case-on-".git" check failed
  from a normal repo subdirectory (where the command returns ../.git, not
  .git or an absolute path), making the prompt resolve to ".." instead of
  the repo name. Applied to ce-compound, ce-sessions, and the
  ce-session-historian agent's Step 1 example.
- extract-metadata.py: defer the full-file --keyword scan until after
  --cwd-filter passes. Previously process_file ran the keyword scan before
  the cwd_filter check, which on Codex (cross-repo discovery) wasted
  scanning on sessions immediately discarded by the filter — recreating
  the long-runtime behavior this work is trying to eliminate.
- extract-metadata.py: emit files_matched: 0 in the empty-input _meta
  branch when --keyword was supplied. Without it, no-result keyword scans
  were ambiguous to the historian's "files_matched: 0 -> stop" rule.
- Tests: added a Codex cross-repo-filtered + --keyword regression test
  and an empty-stdin + --keyword test covering the previously-uncovered
  no-input branch.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

- Reorder Step 3 sub-steps in ce-session-historian.agent.md: drop
  out-of-window sessions and exclude the current session BEFORE applying
  the 5-session deep-dive cap. Previous order capped first, which could
  discard all in-window candidates when high-scoring older sessions
  occupied the cap slots — leaving the agent to falsely return "no
  relevant prior sessions" even when valid in-window matches existed
  further down the candidate list. Tie-breaker rules (branch-match →
  match_count → file size → recency) and STOP semantics unchanged.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>