adapters/claude-code: use CC skill system for inject; split skill telemetry fields (#14)

[lec77]

Linked issue

Closes #14.

Summary

The claude-code adapter's inject mode bypassed CC's native skill system entirely — it dumped the skill content as generic text through --append-system-prompt. CC never registered it as a skill (no init-event listing, no Skill tool dispatch), and the only way SkVM could tell whether the skill was "loaded" was a sentinel-echo heuristic that fails systematically on thinking-mode models. The downstream effect was that DeepSeek-thinking / GLM-thinking jit-optimize runs reported skillLoaded=false, the optimizer abstained at round 1 with infraBlocked, and the entire skill-iteration loop died before producing a single proposal.

This PR rewires inject to use CC's actual skill system (.claude/skills/<name>/SKILL.md + a hard MUST directive forcing Skill tool invocation), and splits the conflated RunResult.skillLoaded field into two semantically distinct signals across all 7 adapters: skillProvided (structural — did we get the skill into the agent's context?) and skillObserved (behavioral — did we see the agent engage?). The mode-aware abstain gate in jit-optimize/workspace.ts now only treats inject-mode skillProvided=false as an infra failure; discover-mode non-loads are merely surfaced as a note.

Changes

Schema & data plumbing (Tasks 1–2):

• core/types.ts: RunResultSchema gains skillProvided?, skillObserved?, skillMode?: "inject" | "discover". skillLoaded retained for one release as a @deprecated mirror so on-disk ~/.skvm/proposals/ evidence files keep parsing.
• jit-optimize/types.ts + jit-optimize/evidence.ts: same three fields land on RunMeta; buildRunMeta forwards them and mirrors skillProvided → skillLoaded (one-way — legacy skillLoaded-only evidence is left alone so older runs aren't retroactively given a skillProvided value we can't actually substantiate).

Optimizer gating (Task 3):

• jit-optimize/workspace.ts:536 — the WARNING line that drives optimizer abstain decisions is now mode-aware:
  • inject + provided=false → WARNING: skill failed to load (inject mode — likely infra issue) (drives abstain as before)
  • discover + provided=false → WARNING: skill not recognized by harness (discover mode — may be description mismatch) (no abstain — discover legitimately allows the model to skip loading)
  • Legacy evidence without skillMode → original WARNING: skill was not loaded wording preserved.
• skillObserved is intentionally NOT surfaced in evidence summaries — it would produce systematic false negatives on thinking-mode models.

claude-code adapter rewrite (Tasks 4–6):

• New buildClaudeCodeInjectArtifacts / buildClaudeCodeDiscoverArtifacts factor the on-disk skill-file work out of run() so it's unit-testable without spawning CC.
• Inject path now writes <workDir>/.claude/skills/<name>/SKILL.md (same path as discover) AND sets --append-system-prompt to a hard MUST directive: You have access to a project skill named `<name>` (<desc>). For this task you MUST invoke the Skill tool with name="<name>" before any other action…
• skillProvided is now read from CC's system.init event's skills[] array — a structural, model-independent signal that resolves the thinking-mode false-negative cleanly. skillObserved is read from Skill/skill tool_use events with matching input.name.
• detectSkillDiscover is split into detectSkillProvided_Discover (init-event) and detectSkillObserved_Discover (Skill tool-use). The old SKILL_INJECT_SENTINEL / injectedSystemPrompt / detectSkillInject machinery is deleted.

Peer adapters — field alignment (Tasks 7–12, one commit each):

• bare-agent, opencode, pi, openclaw, hermes, jiuwenclaw keep their existing inject mechanisms (system-message concat / CONTEXT.md / AGENTS.md / BOOTSTRAP.md / prompt prepend). Each harness's native auto-load file IS the documented inject path; only claude-code had the wrong-mechanism bug.
• All six rewire their skillLoaded plumbing to skillProvided + skillObserved, set skillMode, and write the deprecated skillLoaded mirror. The "agent produced steps → assume skill was used" inference is dropped — skillProvided=true is now set structurally at the inject step site, so the heuristic is redundant. Snippet-match and skill-tool-call detections become skillObserved.
• bare-agent's <load-skill> protocol is the one place where structural and behavioral fire simultaneously — both fields go true together.

Bench report wiring (Task 13):

• bench/types.ts ConditionResult exposes the two new fields; bench/conditions.ts toConditionResult propagates them. skillLoaded deprecated alias retained for older JSON reports.

Test plan

• bunx tsc --noEmit — clean
• bun test — 793 pass / 2 skip / 0 fail (was 777 on origin/main; net +16 tests, no regressions)
• New tests added for: schema field shape, buildRunMeta forwarding (including legacy-only skillLoaded back-compat), mode-aware workspace WARNING wordings, the two new claude-code detection helpers, claude-code inject artifact builders (writes the skill file + emits MUST directive), thinking-mode-safe skillProvided detection from init events alone, bare-agent parallel-assertion coverage.
• Live smoke test against a real model — not yet run. Reviewer or CI can validate against DeepSeek-thinking / GLM-thinking once landed. The integration tests cover the structural contract; the live test is only needed to confirm CC's real CLI accepts the .claude/skills/ + --append-system-prompt directive combination at runtime.

Notes for reviewers

A few design choices worth double-checking:

1. skillMode semantics retained. The inject / discover binary stays — but the meaning is now strictly "loading bias" (forced vs discoverable), independent of delivery mechanism. Each adapter chooses the best mechanism its harness offers. For claude-code, both modes now write to .claude/skills/; the only difference is whether a MUST directive accompanies it. For the other adapters, the existing per-harness inject path (CONTEXT.md / AGENTS.md / etc.) is unchanged because those harnesses don't expose a richer skill subsystem than their auto-load file.

2. Why a system prompt directive and not /skill <name> in the user prompt. The original issue suggested putting /skill <name> in the prompt to trigger the skill. /skill is interactive slash-command syntax — its behavior in claude -p non-interactive mode is undocumented and would have to be verified empirically. System-prompt directives are the documented headless path; they also let us keep the inject/discover semantic distinction crisp (directive is the only difference between the two modes' artifacts).

3. Forward-only mirror for the deprecated skillLoaded. buildRunMeta mirrors skillProvided → skillLoaded but never the reverse: if an older adapter writes only skillLoaded, the new skillProvided on the output stays undefined. Rationale: the old skillLoaded field merged structural and behavioral semantics, so retroactively claiming skillProvided=true from a legacy record would be lying. Consumers (workspace gate) use skillProvided ?? skillLoaded for explicit fallback.

4. Dropped the "agent ran → skill was used" inference in peer adapters. Tests like if (result.steps.length > 0) skillLoaded = true were always weak — they don't tell you anything about the skill, just that the agent produced output. With skillProvided now set structurally at the inject step (Bun.write(...) succeeded), the inference is redundant. Snippet-echo and skill-tool-call detections survive as skillObserved signals (diagnostic only).

5. skillObserved is intentionally NOT surfaced in optimizer evidence summaries. It would systematically false-negative on thinking-mode models (no echo to detect) and create noise. Optimizer code that wants the signal can read runMeta.skillObserved directly.

6. PR scope vs. broader inject redesign. A wider redesign — three modes (forced / discoverable / none) or richer per-harness skill mechanisms — was considered during brainstorming and intentionally deferred. This PR fixes the specific bug in issue [bug] Claude Code adapter inject模式实现Bug报告 #14 and the field-semantics conflation that amplified it. Larger architectural changes can land as a follow-up if useful.

Checklist

• Linked an issue ([bug] Claude Code adapter inject模式实现Bug报告 #14)
• Commit messages follow <scope>: <summary> style
• Added or updated tests where appropriate
• No user-visible CLI flags / on-disk paths changed — no README/docs update needed (the ~/.skvm/proposals/ schema is back-compat via the deprecated skillLoaded alias)
• No unrelated changes mixed in (the 13 commits all serve this one logical change; squash if you prefer fewer atomic commits)

[lec77]

@YYangZiXin Hi! This PR attempts to fix #14, but takes a slightly different approach than what you proposed in the issue — would appreciate your help verifying it.

You suggested putting /skill <name> in the user prompt. The PR instead writes .claude/skills/<name>/SKILL.md and uses --append-system-prompt to add a MUST directive ("You MUST invoke the Skill tool with name=\"<name>\""). Reasoning: /skill is interactive slash-command syntax, and its behavior in claude -p non-interactive mode is undocumented — we didn't want to bet the fix on it.

The PR reads skillProvided from CC's system.init event's skills[] array, assuming CC will auto-discover workDir-scoped .claude/skills/<name>/SKILL.md and list it at init time.

Would you be willing to run the PR branch against your original failing setup (the GLM / DeepSeek thinking mode environment from #14)? Two things to confirm:

1. skillProvided=true appears in the evidence (CC actually recognized the skill)
2. jit-optimize proceeds into Round 1 instead of optimizer abstained (infraBlocked)

git fetch origin pull/20/head:pr-20
git checkout pr-20
# then run the same jit-optimize command that originally hit infra-blocked

Thanks!