chore: promote staging to main (2026-03-10 02:35 UTC)

[ironclaw-ci]

Auto-promotion from staging CI

Batch range: a5f88b32fd0e716df19eaaba4238efc99aa81cc3..83950d11a43ded67b8762b5cf76043d3880f3d7a
Promotion branch: staging-promote/83950d11-22884429853
Base: main
Triggered by: Staging CI batch at 2026-03-10 02:35 UTC

Waiting for gates:

• Tests: pending
• E2E: pending
• Claude Code review: pending (will post comments on this PR)

---

Auto-created by staging-ci workflow

[claude]

Code review

Found 12 issues:

1. [CRITICAL:100] Error handling block unreachable: .await? propagates errors before let Err(msg) pattern can match. Token budget overflow is not logged or marked as failed; job fails silently without diagnostics.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/agent/worker.rs#L494-L502

    if total_tokens > 0
        && let Err(msg) = self
            .context_manager()
            .update_context(self.job_id, |ctx| ctx.add_tokens(total_tokens))
            .await?  // <-- ? propagates error before pattern match
    {
        self.mark_failed(&msg).await?;  // unreachable

2. [CRITICAL:75] TOCTOU race between get_job() and update_job_status() without transaction. Job state can change between check and update; cancellation silently succeeds even if job is no longer active.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/channels/web/handlers/jobs.rs#L283-L305

    let job = store.get_job(job_id).await?;
    if !job.state.is_active() {
        return Err(ApiError::InvalidState(...));
    }
    // <-- Job state can change here
    store.update_job_status(job_id, JobState::Cancelled).await?;

3. [CRITICAL:50] Non-transactional multi-step context updates between metadata/token setup and DB persist. Concurrent worker can modify context between final get_context() and save_job(), causing in-memory/DB divergence. Crash recovery exposes the bug.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/agent/scheduler.rs#L170-L195

    self.context_manager
        .update_context(job_id, |ctx| { ctx.metadata = meta; })
        .await?;
    // ... separate update for token budget ...
    let ctx = self.context_manager.get_context(job_id).await?;
    // <-- Worker can modify ctx here
    store.save_job(&ctx).await?;  // Persists stale snapshot

4. [HIGH:100] Token budget not persisted to database. PR adds max_tokens to JobContext but neither PostgreSQL nor libSQL schemas include the column. Jobs lose token budgets on restart, violating CLAUDE.md dual-backend requirement.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/db/libsql/jobs.rs#L28-L66

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/history/store.rs#L572-L596

    // Neither INSERT nor ON CONFLICT includes max_tokens column
    INSERT INTO agent_jobs (... user_id, ... ) VALUES (... ?16, ... )

5. [HIGH:85] User-supplied metadata bypasses configured token budget without validation. Job metadata can inject {"max_tokens": 9999999} to exceed intended limits.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/agent/scheduler.rs#L164-L168

    let max_tokens = metadata
        .as_ref()
        .and_then(|m| m.get("max_tokens"))
        .and_then(|v| v.as_u64())
        .unwrap_or(self.config.max_tokens_per_job);  // <-- No validation

6. [HIGH:85] Token budget enforcement incomplete: select_tools() LLM calls bypass tracking while respond_with_tools() calls are counted. Jobs can exceed budgets via the planning loop without being caught.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/agent/worker.rs#L489-L501

    // Line 429: select_tools() LLM call — NOT tracked
    let selections = self.select_tools(&mut state, &request).await?;
    // ...
    // Line 460+: Only respond_with_tools() tokens tracked
    let total_tokens = respond_output.usage.total() as u64;
    if total_tokens > 0 && ... ctx.add_tokens(total_tokens) ...

7. [HIGH:75] Unsafe environment variable mutation without synchronization in src/setup/wizard.rs. unsafe { std::env::set_var() } and remove_var() mutate global state; concurrent tests cause data races.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/setup/wizard.rs#L3661-L3676

    unsafe {
        std::env::set_var(key, value);  // <-- No locking, test parallel race
    }

8. [HIGH:75] Removed error handling guard in .github/workflows/staging-ci.yml. Changed from if: ${{ secrets.GH_RELEASES_MANAGER_APP_ID != '' }} to always attempting token generation. Workflow fails hard instead of gracefully degrading to fallback token.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/.github/workflows/staging-ci.yml#L69

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/.github/workflows/staging-ci.yml#L230

    - name: Create GitHub App token
      id: app-token
      uses: actions/create-github-app-token@v1
      // <-- Previously had: if: ${{ secrets.GH_RELEASES_MANAGER_APP_ID != '' }}

9. [HIGH:75] Breaking UX change: tool_remove and skill_remove downgraded from auto-approvable to always-require-approval without release notes, affecting workflows where users previously auto-approved these tools.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/tools/builtin/extension_tools.rs#L494-L495

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/tools/builtin/skill_tools.rs#L911-L912

    approval: ApprovalRequirement::Always,  // Changed from UnlessAutoApproved

10. [MEDIUM:75] Stringly-typed token budget with semantic magic number 0 = unlimited violates stated code style preference for strong types. Requires defensive > 0 checks in 3+ places; error-prone for future maintainers.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/config/agent.rs#L485

    max_tokens_per_job: u64,  // Semantic: 0 = unlimited, documented only in comments
    // Better: enum TokenBudget { Unlimited | Limited(u64) }

11. [MEDIUM:75] Missing error context mapping in token budget enforcement. Error is bare String instead of structured JobError::ContextError. Loss of context about which LLM phase (respond vs select) caused overflow makes debugging harder.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/agent/worker.rs#L499-L500

    self.mark_failed(&msg).await?;  // msg is bare String, no phase context

12. [LOW:50] Configuration precedence for max_tokens_per_job not documented in one place. Resolution order spans job metadata, env var, settings.json, and default—correct implementation but non-obvious to maintainers and operators.

https://github.com/anthropics/ironclaw/blob/577e26eff4960cd31932e22ef017ae1599f3fbc6/src/agent/scheduler.rs#L163-L168