Sigrid Jin relayed an adamantium Discord field report: Anthropic rejected requests with invalid_request_error when messages contained tool_use ids without immediately following tool_result blocks.
Coalesce consecutive tool-result messages after assistant tool_use blocks into one Anthropic user message, and drop orphan tool_use/tool_result blocks before dispatch so resume/edit/compaction boundary damage cannot reach the provider as a 400.
Tests cover parallel tool results and orphaned resume-boundary history.
Per gaebal-gajae cycle #105 validation pass. One-liner state summary
now appears at top (tone-setter for reviewers) and bottom (reinforced
recap):
'Phase 0 is now frozen, reviewer-mapped, and merge-ready;
Phase 1 remains intentionally deferred behind the locked priority order.'
This is the single authoritative sentence that captures branch state.
Use it for PR titles, review summaries, and Phase 1 handoff notes.
Why this framing matters (per gaebal-gajae evaluation):
- 'frozen' signals no scope creep
- 'reviewer-mapped' signals audit trail exists (this guide)
- 'merge-ready' signals gates are passed
- 'intentionally deferred' signals Phase 1 absence is by design, not omission
- 'locked priority order' signals sequencing is validated (cycle #104-#105)
Review guide now doubles as merge-enabler: reviewers parse branch state
in one sentence, then drill into commits as needed.
Doc-only. No code changes. Freeze preserved.
Pre-merge documentation for reviewers. Summarizes:
- What Phase 0 tasks deliver (JSON envelope contracts, regression locks)
- Why dogfood cycles #99-#105 matter (validated methodology, 15 filed pinpoints)
- Commit-by-commit navigation for the 30-commit frozen bundle
- What lands vs what's deferred
- Integration notes for Phase 1 planning
- Known limitations + follow-ups
This is doc-only, no code changes. Serves as audit trail and reviewer
reference without adding scope to the frozen feature branch.
Pinpoint #172: SCHEMAS.md v1.5 Emission Baseline documentation inaccuracy
discovered during cycle #98 probe.
The Phase 1 normalization targets section claimed:
"unify where `action` field appears (only in 4 inventory verbs)"
But reality is only 3 inventory verbs have `action`:
- mcp
- skills
- agents
list-sessions uses `command` instead (the documented 1-of-13 deviation
already captured elsewhere in v1.5 baseline).
This is a doc-truthfulness issue (same family as cycles #76, #79, #82).
Active misdocumentation leads downstream consumers to assume 4-verb
coverage when building adapters/dispatchers.
Changes:
1. SCHEMAS.md: 'only in 4 inventory verbs' → 'only in 3 inventory verbs: mcp, skills, agents'
2. Added regression test `v1_5_action_field_appears_only_in_3_inventory_verbs_172`
- Asserts mcp/skills/agents HAVE action field
- Asserts help/version/doctor/status/sandbox/system-prompt/bootstrap-plan/list-sessions do NOT have action field
- Forces SCHEMAS.md + binary to stay synchronized
Test added:
- `v1_5_action_field_appears_only_in_3_inventory_verbs_172` (8 negative cases + 3 positive cases)
Tests: 227/227 pass (+1 from #172).
Related: #155 (doc parity family), #168c (emission baseline).
Doc-truthfulness family: #76, #79, #82, #172.
Pinpoint #171: typed-error classifier gap discovered during #141 probe cycle #97.
`claw list-sessions --help` emits:
error: unexpected extra arguments after `claw list-sessions`: --help
This format is used by multiple verbs that reject trailing positional args:
- list-sessions
- plugins (subcommands)
- config (subcommands)
- diff
- load-session
Before fix:
{"error": "unexpected extra arguments after `claw list-sessions`: --help",
"hint": null,
"kind": "unknown",
"type": "error"}
After fix:
{"error": "unexpected extra arguments after `claw list-sessions`: --help",
"hint": "Run `claw --help` for usage.",
"kind": "cli_parse",
"type": "error"}
The pattern `unexpected extra arguments after \`claw` is specific enough
that it won't hijack generic prose mentioning "unexpected extra arguments"
in other contexts (sanity test included).
Side benefit: like #169/#170, correctly classified cli_parse errors now
auto-trigger the #247 hint synthesizer.
Related #141 gap not yet closed: `claw list-sessions --help` still errors
instead of showing help (requires separate parser fix to recognize --help
as a distinct path). This classifier fix at least makes the error surface
typed correctly so consumers can distinguish "parse failure" from "unknown"
and potentially retry without the --help flag.
Test added:
- `classify_error_kind_covers_unexpected_extra_args_171` (4 positive cases
+ 1 sanity guard)
Tests: 226/226 pass (+1 from #171).
Typed-error family: #121, #127, #129, #130, #164, #169, #170, #247.
Pinpoint #153 closure. USAGE.md was missing practical instructions for:
1. Adding the claw binary to PATH (symlink vs export PATH)
2. Verifying the install works (version, doctor, --help)
3. Troubleshooting PATH issues (which, echo $PATH, ls -la)
New subsections:
- "Add binary to PATH" with two common options
- "Verify install" with post-install health checks
- Troubleshooting guide for common failures
Target audience: developers building from source who want to run `claw`
from any directory without typing `./rust/target/debug/claw`.
Discovered during cycle #96 dogfood (10-min reminder cycle).
Tests: 225/225 still pass (doc-only change).
Pinpoint #170: Extended typed-error classifier coverage gap discovered during
dogfood probe 2026-04-23 07:30 Seoul (cycle #95).
The #169 comment claimed to cover `--permission-mode bogus` via the
`unsupported value for --` pattern, but the actual `parse_permission_mode_arg`
message format is `unsupported permission mode 'bogus'` (NO `for --` prefix).
Doc-vs-reality lie in the #169 fix itself — fixed here.
Four classifier gaps closed:
1. `unsupported permission mode '<value>'` → cli_parse
(from: `parse_permission_mode_arg`)
2. `invalid value for --reasoning-effort: '<value>'; must be ...` → cli_parse
(from: `--reasoning-effort` validator)
3. `model string cannot be empty` → cli_parse
(from: empty --model rejection)
4. `slash command /<name> is interactive-only. Start \`claw\` ...` →
slash_command_requires_repl (NEW kind — more specific than cli_parse)
The fourth pattern gets its own kind (`slash_command_requires_repl`) because
it's a command-mode misuse, not a parse error. Downstream consumers can
programmatically offer REPL-launch guidance.
Side benefit: like #169, the correctly classified cli_parse errors now
auto-trigger the #247 hint synthesizer ("Run `claw --help` for usage.").
Test added:
- `classify_error_kind_covers_flag_value_parse_errors_170_extended`
(4 positive cases + 2 sanity guards)
Tests: 225/225 pass (+1 from #170).
Typed-error family: #121, #127, #129, #130, #164, #169, #247.
Discovered via systematic probe angle: 'error message pattern audit' \u2014
grep each error emission for pattern, confirm classifier matches.
Pinpoint #169: typed-error classifier gap discovered during dogfood probe.
`claw --output-format json --output-format xml doctor` was emitting:
{"error": "unsupported value for --output-format: xml ...",
"hint": null,
"kind": "unknown",
"type": "error"}
After fix:
{"error": "unsupported value for --output-format: xml ...",
"hint": "Run `claw --help` for usage.",
"kind": "cli_parse",
"type": "error"}
The change adds two new classifier branches to `classify_error_kind`:
1. `unsupported value for --` → cli_parse
2. `missing value for --` → cli_parse
Covers all `CliOutputFormat::parse` / `parse_permission_mode_arg` rejections
and any future flag-value validation messages using the same pattern.
Side benefit: the #247 hint synthesizer ("Run `claw --help` for usage.")
now triggers automatically because the error is now correctly classified
as cli_parse. Consumers get both correct kind AND helpful hint.
Test added:
- `classify_error_kind_covers_flag_value_parse_errors_169` (4 positive +
1 sanity case)
Tests: 224/224 pass (+1 from #169).
Discovered during dogfood probe 2026-04-23 07:00 Seoul, cycle #94.
Refs: #169, typed-error family (#121, #127, #129, #130, #164, #247)
Pinpoint #155: USAGE.md was missing documentation for three interactive
commands that appear in `claw --help`:
- /ultraplan [task]
- /teleport <symbol-or-path>
- /bughunter [scope]
Also adds full documentation for other underdocumented commands:
- /commit, /pr, /issue, /diff, /plugin, /agents
Converts inline sentence list into structured section 'Interactive slash
commands (inside the REPL)' with brief descriptions for each command.
Closes#155 gap: discovered during dogfood probing of help/USAGE parity.
No code changes. Pure documentation update.
Phase 0 Task 4 of the JSON Productization Program: CI shape parity guard.
This test locks the v1.5 emission baseline (documented in SCHEMAS.md § v1.5
Emission Baseline) so any future PR that introduces shape drift in a documented
verb fails this test at PR time.
Complements Task 2 (no-silent guarantee) by asserting SPECIFIC top-level key
sets, not just 'stdout is non-empty valid JSON'. If a verb adds/removes a
top-level field, this test fails with a clear error message pointing to
SCHEMAS.md § v1.5 Emission Baseline for update guidance.
Coverage:
- 8 success-path verbs with locked shape (help, version, doctor, skills,
agents, system-prompt, bootstrap-plan, list-sessions)
- 2 error-path cases with locked error envelope shape (prompt-no-arg, doctor --foo)
Key enforcement rules:
- Success envelope: exact key set match per verb
- Error envelope: {error, hint, kind, type} (4 keys, all verbs)
- list-sessions deliberately kept as {command, sessions} (Phase 1 target)
Test design intent:
- Locks CURRENT (possibly imperfect) shape, NOT target shape
- Forces PR authors to update both code + SCHEMAS.md + test together
- Makes Phase 1 shape normalization PRs visible: 'update this test'
Phase 0 now COMPLETE:
- Task 1 ✅ Stream routing fix (cycle #89)
- Task 2 ✅ No-silent guarantee (cycle #90)
- Task 3 ✅ Per-verb emission inventory SCHEMAS.md (cycle #91)
- Task 4 ✅ CI shape parity guard (this cycle)
Tests: 18 output_format_contract tests all pass (+1 from Task 4).
v1.5 emission baseline now locked by code + tests + docs.
Refs: #168c, cycle #92, Phase 0 Task 4 (final)
Under --output-format json, error envelopes were emitted to stderr via
eprintln!. This violated the emission contract: stdout should carry the
contractual envelope (success OR error); stderr is reserved for
non-contractual diagnostics.
Cycle #87 controlled matrix audit found bootstrap/dump-manifests/state
exhibited this pattern (exit 1, stdout 0 bytes, stderr N bytes under
--output-format json).
Fix: change eprintln! to println! for the JSON error envelope path in main().
Text mode continues to route errors to stderr (conventional).
Verification:
- bootstrap --output-format json: stdout now carries envelope, exit 1
- dump-manifests --output-format json: stdout now carries envelope, exit 1
- Text mode: errors still on stderr with [error-kind: ...] prefix (no regression)
Tests:
- Updated assert_json_error_envelope helper to read from stdout (was stderr)
- Added error_envelope_emitted_to_stdout_under_output_format_json_168c
regression test that asserts envelope on stdout + non-JSON on stderr
- All 16 output_format_contract tests pass
Phase 0 Task 1 complete: emission routing fixed across all error-path verbs.
Phase 0 Task 2 (no-silent CI guarantee) remains.
Refs: #168c (cycle #87 filing), cycle #88 emission contract framing
Fresh-dogfood validation (cycle #84, #168) proved the original locus premise was
underspecified. v1.0 was never a coherent contract — each verb has a bespoke JSON
shape with no coordination, and bootstrap JSON is completely broken (silent
failure, exit 0 no output).
Revised migration plan:
- Phase 0 (NEW): Emergency fix for silent failures (#168 bootstrap JSON)
- Phase 1 (NEW): v1.5 baseline — minimal JSON invariants across all 14 verbs
- Every command emits valid JSON with --output-format json
- Every command has top-level 'kind' field for verb ID
- Every error envelope follows {error, hint, kind, type}
- Phase 2 (renamed from Phase 1): v2.0 wrapped envelope (opt-in)
- Phase 3 (renamed from Phase 2): v2.0 default
- Phase 4 (renamed from Phase 3): v1.0/v1.5 deprecation
Rationale:
- Can't migrate from 'incoherent' to 'coherent v2.0' in one jump
- Consumers need stable target (v1.5) to transition from
- Silent failures must be fixed BEFORE migration (consumers can't detect breakage)
Effort revision: ~9 dev-days (Phase 0: 1 + Phase 1: 3 + Phase 2: 5) vs original
~6 dev-days for direct v1.0→v2.0 (which would have failed).
Doctrine implication: Fresh-dogfood principle (#9, cycle #73) prevented a multi-day
migration from hitting an unsolvable baseline problem. Evidence-backed mid-design
correction.
SCHEMAS.md was presenting the target v2.0 schema as the current binary contract.
This is the source of truth document, so the misdocumentation propagated to every
downstream doc (USAGE.md, ERROR_HANDLING.md, CLAUDE.md all inherited the false
premise that v1.0 includes timestamp/command/exit_code/etc).
Fixed with:
1. CRITICAL header at top: marks entire doc as v2.0 target, not v1.0 reality
2. 'TARGET v2.0 SCHEMA' headers on Common Fields section
3. Comprehensive Appendix: v1.0 actual shape + migration timeline + v1.0 code example
4. Links to FIX_LOCUS_164.md + ERROR_HANDLING.md for v1.0 reality
5. FAQ: clarifies the version mismatch and when v2.0 ships
This closes the fourth P0 doc-truthfulness instance (4/4 in family):
- #78 USAGE.md: active misdocumentation (fixed#78)
- #79 ERROR_HANDLING.md: copy-paste trap (fixed#79)
- #165 CLAUDE.md: boundary collapse (fixed#81)
- #166 SCHEMAS.md: aspirational source doc (fixed#82)
Pattern is now crystallized: SCHEMAS.md was the aspirational source;
three downstream docs (USAGE, ERROR_HANDLING, CLAUDE) inherited the false v2.0-as-v1.0
claim. Fix the source (SCHEMAS.md), which eliminates the root cause for all four.
CLAUDE.md was documenting the v2.0 target schema as if it were current binary
behavior. This misled validator/harness implementers into assuming the Rust
binary emits timestamp, command, exit_code, output_format, schema_version fields
when it doesn't.
Fixed by explicitly marking the boundary:
1. SCHEMAS.md section: now clearly labels 'target v2.0 design' and lists both
v1.0 (actual binary) and v2.0 (target) field shapes
2. Clawable commands requirements: now explicitly separates v1.0 (current) and
v2.0 (post-FIX_LOCUS_164) envelope requirements
3. Added inline migration note pointing to FIX_LOCUS_164.md
This closes#165 as the third P0 doc-truthfulness fix (Option A: preserve current
truth, add v2.0 target as separate labeled section).
P0 doc-truthfulness family pattern (all three related to #164 envelope divergence):
- #78 USAGE.md: active misdocumentation (fixed cycle #78)
- #79 ERROR_HANDLING.md: copy-paste trap (fixed cycle #79)
- #165 CLAUDE.md: target/current boundary collapse (fixed cycle #81)
The Python code examples were accessing nested error.kind like envelope['error']['kind'],
but v1.0 emits flat envelopes with error as a STRING and kind at top-level.
Updated:
- Table header: now shows actual v1.0 shape {error: "...", kind: "...", type: "error"}
- match statement: switched from envelope.get('error',{}).get('kind') to envelope.get('kind')
- All ClawError raises: changed from envelope['error']['message'] to envelope.get('error','')
because error field is a STRING in v1.0, not a nested object
- Added inline comments on every error case noting v1.0 vs v2.0 difference
- Appendix: split into v1.0 (actual/current) and v2.0 (target after FIX_LOCUS_164)
The code examples now work correctly against the actual binary.
This was active misdocumentation (P0 severity) — the Python examples would crash
if a consumer tried to use them.
The JSON output section was misleading — it claimed the binary emits
exit_code, command, timestamp, output_format, schema_version, and nested
error objects. The binary actually emits v1.0 flat shape (kind at top-level,
error as string, no common metadata fields).
Updated section:
- Documents actual v1.0 success and error envelope shapes
- Lists known issues (missing fields, overloaded kind, flat error)
- Shows how to dispatch on v1.0 (check type=='error' before reading kind)
- Warns users NOT to rely on kind alone
- Links to FIX_LOCUS_164.md for migration plan
- Explains Phase 1/2/3 timeline for v2.0 adoption
This is a doc-only fix that makes USAGE.md truthful about the current behavior
while preparing users for the coming schema migration.
Problem: In git worktrees, .git is a pointer file (not a directory), so cargo's
rerun-if-changed=.git/HEAD never triggers when commits are made. This causes
claw version to report a stale SHA after new commits.
Solution: Add resolve_git_head_path() helper that detects worktree mode:
- If .git is a file: parse gitdir pointer, watch <gitdir>/HEAD
- If .git is a directory: watch .git/HEAD (regular repo)
This ensures build.rs invalidates on each commit, making version output truthful.
Verification: Binary built in worktree now reports correct SHA after commits
(before: stale, after: current HEAD).
Relates to ROADMAP #161 (filed cycle #65, implemented cycle #69).
Diagnostic-strictness family member.
Diff: 21 lines added (resolve_git_head_path + conditional rerun-if-changed).
