Cycle #27 ships a new test class systematizing the three-layer protocol
invariant framework.
## Context
After cycles #20–#26, the protocol has three distinct invariant classes:
1. **Structural compliance** (#178): Does the envelope exist?
2. **Quality compliance** (#179): Is stderr silent + error message truthful?
3. **Cross-channel consistency** (#181 + NEW): Do multiple channels agree?
#181 revealed a critical gap: the second test class was incomplete.
Envelopes could be structurally valid, quality-compliant, but still
lie about their own state (envelope.exit_code != actual exit).
## New Test Class
TestCrossChannelConsistency in test_cross_channel_consistency.py captures
the third invariant layer with 5 dedicated tests:
1. envelope.command ↔ dispatched subcommand
2. envelope.output_format ↔ --output-format flag
3. envelope.timestamp ↔ actual wall clock (recent, <5s)
4. envelope.exit_code ↔ process exit code (cycle #26/#181 regression guard)
5. envelope boolean fields (found/handled/deleted) ↔ error block presence
Each test specifically targets cross-channel truth, not structure or quality.
## Why Separate Test Classes Matter
A command can fail all three ways independently:
| Failure mode | Exit/Crash | Test class | Example |
|---|---|---|---|
| Structural | stderr noise | TestParseErrorEnvelope | argparse leaks to stderr |
| Quality | correct shape, wrong message | TestParseErrorStderrHygiene | error instead of real message |
| Cross-channel | truthy field, lie about state | TestCrossChannelConsistency | exit_code: 0 but exit 1 |
#181 was invisible to the first two classes. A claw passing all structure/
quality tests could still be misled. The third class catches that.
## Audit Results (Cycle #27)
All 5 tests pass — no drift detected in any channel pair:
- ✅ Envelope command always matches dispatch
- ✅ Envelope output_format always matches flag
- ✅ Envelope timestamp always recent (<5s)
- ✅ Envelope exit_code always matches process exit (post-#181 guard)
- ✅ Boolean fields consistent with error block presence
The systematic audit proved the fix from #181 holds, and identified
no new cross-channel gaps.
## Test Impact
- 209 → 214 tests passing (+5)
- Zero regressions
- New invariant class now has dedicated test suite
- Future cross-channel bugs will be caught by this class
## Related
- #178 (#20): Parser-front-door structural contract
- #179 (#20): Stderr hygiene + real error message quality
- #181 (#26): Envelope exit_code must match process exit
- #182-N: Future cross-channel contract violations will be caught
by TestCrossChannelConsistency
This test class is evergreen — as new fields/channels are added to the
protocol, invariants for those channels should be added here, not mixed
with other test classes. Keeping invariant classes separate makes
regression attribution instant (e.g., 'TestCrossChannelConsistency failed'
= 'some truth channel disagreed').
Classification (per cycle #24 calibration):
- Red-state bug: ✗ (audit is green)
- Real friction: ✓ (structured audit of documented invariants)
- Proof of equilibrium: ✓ (systematic verification, no gaps found)
Source: Jobdori cycle #27 proactive invariant audit — following gaebal
guidance to probe documented invariants, not speculative gaps.
Cycle #26 dogfood found a real red-state bug in the JSON envelope contract.
## The Bug
exec-command and exec-tool not-found cases return exit code 1 from the
process, but the envelope reports exit_code: 0 (the default from
wrap_json_envelope). This is a protocol violation.
Repro (before fix):
$ claw exec-command unknown-cmd test --output-format json > out.json
$ echo $?
1
$ jq '.exit_code' out.json
0 # WRONG — envelope lies about exit code
Claws reading the envelope's exit_code field get misinformation. A claw
implementing the canonical ERROR_HANDLING.md pattern (check exit_code,
then classify by error.kind) would incorrectly treat failures as
successes when dispatching on the envelope alone.
## Root Cause
main.py lines 687–739 (exec-command + exec-tool handlers):
- Return statement: 'return 0 if result.handled else 1' (correct)
- Envelope wrap: 'wrap_json_envelope(envelope, args.command)'
(uses default exit_code=0, IGNORES the return value)
The envelope wrap was called BEFORE the return value was computed, so
the exit_code field was never synchronized with the actual exit code.
## The Fix
Compute exit_code ONCE at the top:
exit_code = 0 if result.handled else 1
Pass it explicitly to wrap_json_envelope:
wrap_json_envelope(envelope, args.command, exit_code=exit_code)
Return the same value:
return exit_code
This ensures the envelope's exit_code field is always truth — the SAME
value the process returns.
## Tests Added (3)
TestEnvelopeExitCodeMatchesProcessExit in test_exec_route_bootstrap_output_format.py:
1. test_exec_command_not_found_envelope_exit_matches:
Verifies exec-command unknown-cmd returns exit 1 in both envelope
and process.
2. test_exec_tool_not_found_envelope_exit_matches:
Same for exec-tool.
3. test_all_commands_exit_code_invariant:
Audit across 4 known non-zero cases (show-command, show-tool,
exec-command, exec-tool not-found). Guards against the same bug
in other surfaces.
## Impact
- 206 → 209 passing tests (+3)
- Zero regressions
- Protocol contract now truthful: envelope.exit_code == process exit
- Claws using the one-handler pattern from ERROR_HANDLING.md now get
correct information
## Related
- ERROR_HANDLING.md (cycle #22): Documented exit_code as machine-readable
contract field
- #178/#179 (cycles #19/#20): Closed parser-front-door contract
- This closes a gap in the WORK PROTOCOL contract — envelope values must
match reality, not just be structurally present.
Classification (per cycle #24 calibration):
- Red-state bug: ✓ (contract violation, claws get misinformation)
- Real friction: ✓ (discovered via dogfood, not speculative)
- Fix ships same-cycle: ✓ (discipline per maintainership mode)
Source: Jobdori cycle #26 dogfood — ran multiple edge-case probes, noticed
exec-command envelope showed exit_code: 0 while process exited 1.
Investigated wrap_json_envelope default behavior, confirmed bug, fixed
and tested in same cycle.
Dogfood discovered #178 had two residual gaps:
1. Stderr pollution: argparse usage + error text still leaked to stderr even in
JSON mode (envelope was correct on stdout, but stderr noise broke the
'machine-first protocol' contract — claws capturing both streams got dual output)
2. Generic error message: envelope carried 'invalid command or argument (argparse
rejection)' instead of argparse's actual text like 'the following arguments
are required: session_id' or 'invalid choice: typo (choose from ...)'
Before #179:
$ claw load-session --output-format json
[stdout] {"error": {"message": "invalid command or argument (argparse rejection)"}}
[stderr] usage: main.py load-session [-h] ...
main.py load-session: error: the following arguments are required: session_id
[exit 1]
After #179:
$ claw load-session --output-format json
[stdout] {"error": {"message": "the following arguments are required: session_id"}}
[stderr] (empty)
[exit 1]
Implementation:
- New _ArgparseError exception class captures argparse's real message
- main() monkey-patches parser.error (+ all subparser.error) in JSON mode to raise
_ArgparseError instead of print-to-stderr + sys.exit(2)
- _emit_parse_error_envelope() now receives the real message verbatim
- Text mode path unchanged: still uses original argparse print+exit behavior
Contract:
- JSON mode: stdout carries envelope with argparse's actual error; stderr silent
- Text mode: unchanged — argparse usage to stderr, exit 2
- Parse errors still error.kind='parse', retryable=false
Test additions (5 new, 14 total in test_parse_error_envelope.py):
- TestParseErrorStderrHygiene (5):
- test_json_mode_stderr_is_silent_on_unknown_command
- test_json_mode_stderr_is_silent_on_missing_arg
- test_json_mode_envelope_carries_real_argparse_message
- test_json_mode_envelope_carries_invalid_choice_details (verifies valid-choices list)
- test_text_mode_stderr_preserved_on_unknown_command (backward compat)
Operational impact:
Claws capturing both stdout and stderr no longer get garbled output. The envelope
message now carries discoverability info (valid command list, missing-arg name)
that claws can use for retry/recovery without probing the CLI a second time.
Test results: 201 → 206 passing, 3 skipped unchanged, zero regression.
Pinpoint discovered via dogfood at 2026-04-22 20:30 KST (cycle #20).
Adds parametrised test suite validating that clawable-surface commands'
JSON output matches their declared envelope contracts per SCHEMAS.md.
Two phases:
Phase 1 (this commit): Consistency baseline.
- Collect ENVELOPE_CONTRACTS registry mapping each command to its
required and optional fields
- TestJsonEnvelopeConsistency: parametrised test iterates over 13
commands, invokes with --output-format json, validates that
actual JSON envelope contains all required fields
- test_envelope_field_value_types: spot-check types (int, str, list)
for consistency
Phase 2 (future #173): Common field wrapping.
- Once wrap_json_envelope() is applied, all commands will emit
timestamp, command, exit_code, output_format, schema_version
- Currently skipped via @pytest.mark.skip, these tests will activate
automatically when wrapping is implemented:
TestJsonEnvelopeCommonFieldPrep::test_all_envelopes_include_timestamp
TestJsonEnvelopeCommonFieldPrep::test_all_envelopes_include_command
TestJsonEnvelopeCommonFieldPrep::test_all_envelopes_include_exit_code_and_schema_version
Why this matters:
- #172 documented the JSON contract; this test validates it
- Currently detects when actual output diverges from SCHEMAS.md
(e.g. list-sessions emits 'count', not 'sessions_count')
- As #173 wraps commands, test suite auto-validates new common fields
- Prevents regression: accidental field removal breaks the test suite
Current status: 11 passed (consistency), 6 skipped (awaiting #173)
Full suite: 168 → 179 passing, zero regression.
Closes ROADMAP #173 prep (framework for common field validation).
Actual field wrapping remains for next cycle.
Stops manual parity inspection from being a human-noticed concern. When
a developer adds a new subcommand to the claw-code CLI, this test suite
enforces explicit classification:
- CLAWABLE_SURFACES: MUST accept --output-format {text,json}
- OPT_OUT_SURFACES: explicitly exempt with documented rationale
A new command that forgets to opt into one of these two sets FAILS
loudly with TestCommandClassificationCoverage::test_every_registered_
command_is_classified. No silent drift possible.
Technique: argparse introspection at test time walks the _actions tree,
discovers every registered subcommand, and compares against the declared
classification sets. Contract is enforced machine-first instead of
depending on human review.
Three test classes covering three invariants:
TestClawableSurfaceParity (14 tests):
- test_all_clawable_surfaces_accept_output_format: every member of
CLAWABLE_SURFACES has --output-format flag registered
- test_clawable_surface_output_format_choices (parametrised over 13
commands): each must accept exactly {text, json} and default to 'text'
for backward compat
TestCommandClassificationCoverage (3 tests):
- test_every_registered_command_is_classified: any new subcommand
must be explicitly added to CLAWABLE_SURFACES or OPT_OUT_SURFACES
- test_no_command_in_both_sets: sanity check for classification conflicts
- test_all_classified_commands_actually_exist: no phantom commands
(catches stale entries after a command is removed)
TestJsonOutputContractEndToEnd (10 tests):
- test_command_emits_parseable_json (parametrised over 10 clawable
commands): actual subprocess invocation with --output-format json
produces valid parseable JSON on stdout
Classification:
CLAWABLE_SURFACES (13):
Session lifecycle: list-sessions, delete-session, load-session,
flush-transcript
Inspect: show-command, show-tool
Execution: exec-command, exec-tool, route, bootstrap
Diagnostic inventory: command-graph, tool-pool, bootstrap-graph
OPT_OUT_SURFACES (12):
Rich-Markdown reports (future JSON schema): summary, manifest,
parity-audit, setup-report
List filter commands: subsystems, commands, tools
Turn-loop: structured_output is future work
Simulation/debug: remote-mode, ssh-mode, teleport-mode,
direct-connect-mode, deep-link-mode
Full suite: 141 → 168 passing (+27), zero regression.
Closes ROADMAP #171.
Why this matters:
Before: parity was human-monitored; every new command was a drift
risk. The CLUSTER 3 sweep required manually auditing every
subcommand and landing fixes as separate pinpoints.
After: parity is machine-enforced. If a future developer adds a new
command without --output-format, the test suite blocks it
immediately with a concrete error message pointing at the
missing flag.
This is the first step in Gaebal-gajae's identified upper-level work:
operationalised parity instead of aspirational parity.
Related clusters:
- Clawability principle: machine-first protocol enforcement
- Test-first regression guard: extends TestTripletParityConsistency
(#160/#165) and TestFullFamilyParity (#166) from per-cluster
parity to cross-surface parity
Extends the diagnostic surface audit with the two inventory-structure
commands: command-graph (command family segmentation) and tool-pool
(assembled tool inventory). Both now expose their underlying rich
datastructures via JSON envelope.
Concrete additions:
- command-graph: --output-format {text,json}
- tool-pool: --output-format {text,json}
JSON envelope shapes:
command-graph:
{builtins_count, plugin_like_count, skill_like_count, total_count,
builtins: [{name, source_hint}],
plugin_like: [{name, source_hint}],
skill_like: [{name, source_hint}]}
tool-pool:
{simple_mode, include_mcp, tool_count,
tools: [{name, source_hint}]}
Backward compatibility:
- Default is 'text' (Markdown unchanged)
- Text output byte-identical to pre-#169
Tests (4 new, test_command_graph_tool_pool_output_format.py):
- TestCommandGraphOutputFormat (2): JSON structure + text compat
- TestToolPoolOutputFormat (2): JSON structure + text compat
Full suite: 137 → 141 passing, zero regression.
Closes ROADMAP #169.
Why this matters:
Claws auditing the codebase can now ask 'what commands exist' and
'what tools exist' and get structured, parseable answers instead of
regex-parsing Markdown headers and counting list items.
Related clusters:
- Diagnostic surfaces (#169 adds to #167/#168 work-verb parity)
- Inventory introspection (command-graph + tool-pool are the two
foundational 'what do we have?' queries)
Closes the inspect-capability parity gap: show-command and show-tool were
the only discovery/inspection CLI commands lacking --output-format support,
making them outliers in the ecosystem that already had unified JSON
contracts across list-sessions, load-session, delete-session, and
flush-transcript (#160/#165/#166).
Concrete additions:
- show-command: --output-format {text,json}
- show-tool: --output-format {text,json}
JSON envelope shape (found case):
{name, found: true, source_hint, responsibility}
JSON envelope shape (not-found case):
{name, found: false, error: {kind:'command_not_found'|'tool_not_found',
message, retryable: false}}
Exit codes:
0 = success
1 = not found
Backward compatibility:
- Default (no --output-format) is 'text' (unchanged)
- Text output byte-identical to pre-#167 (three newline-separated lines)
Tests (10 new, test_show_command_tool_output_format.py):
- TestShowCommandOutputFormat (5): found + not-found in JSON; text mode
backward compat; text is default
- TestShowToolOutputFormat (3): found + not-found in JSON; text mode
backward compat
- TestShowCommandToolFormatParity (2): both accept same flag choices;
consistent JSON envelope shape
Full suite: 114 → 124 passing, zero regression.
Closes ROADMAP #167.
Why this matters:
Before: Claws calling show-command/show-tool had to parse human-readable
prose output via regex, with no structured error signal.
After: Same envelope contract as load-session and friends: JSON-first,
typed errors, machine-parseable.
Related clusters:
- Session-lifecycle CLI parity family (#160, #165, #166, #167)
- Machine-readable error contracts (same vein as #162 atomicity + #164
cancellation state-safety: structured boundaries for orchestration)
Closes the #161 follow-up gap identified in review: wall-clock timeout
bounded caller-facing wait but did not cancel the underlying provider
thread, which could silently mutate mutable_messages / transcript_store /
permission_denials / total_usage after the caller had already observed
stop_reason='timeout'. A ghost turn committed post-deadline would poison
any session that got persisted afterwards.
Stage A scope (this commit): runtime + engine layer cooperative cancel.
Engine layer (src/query_engine.py):
- submit_message now accepts cancel_event: threading.Event | None = None
- Two safe checkpoints:
1. Entry (before max_turns / budget projection) — earliest possible return
2. Post-budget (after output synthesis, before mutation) — catches cancel
that arrives while output was being computed
- Both checkpoints return stop_reason='cancelled' with state UNCHANGED
(mutable_messages, transcript_store, permission_denials, total_usage
all preserved exactly as on entry)
- cancel_event=None preserves legacy behaviour with zero overhead (no
checkpoint checks at all)
Runtime layer (src/runtime.py):
- run_turn_loop creates one cancel_event per invocation when a deadline
is in play (and None otherwise, preserving legacy fast path)
- Passes the same event to every submit_message call across turns, so a
late cancel on turn N-1 affects turn N
- On timeout (either pre-call or mid-call), runtime explicitly calls
cancel_event.set() before future.cancel() + synthesizing the timeout
TurnResult. This upgrades #161's best-effort future.cancel() (which
only cancels not-yet-started futures) to cooperative mid-flight cancel.
Stop reason taxonomy after Stage A:
'completed' — turn committed, state mutated exactly once
'max_budget_reached' — overflow, state unchanged (#162)
'max_turns_reached' — capacity exceeded, state unchanged
'cancelled' — cancel_event observed, state unchanged (#164 Stage A)
'timeout' — synthesised by runtime, not engine (#161)
The 'cancelled' vs 'timeout' split matters:
- 'timeout' is the runtime's best-effort signal to the caller: deadline hit
- 'cancelled' is the engine's confirmation: cancel was observed + honoured
If the provider call wedges entirely (never reaches a checkpoint), the
caller still sees 'timeout' and the thread is leaked — but any NEXT
submit_message call on the same engine observes the event at entry and
returns 'cancelled' immediately, preventing ghost-turn accumulation.
This is the honest cooperative limit in Python threading land; true
preemption requires async-native provider IO (future work, not Stage A).
Tests (29 new tests, tests/test_submit_message_cancellation.py + tests/
test_run_turn_loop_cancellation.py):
Engine-layer (12 tests):
- TestCancellationBeforeCall (5): pre-set event returns 'cancelled' immediately;
mutable_messages, transcript_store, usage, permission_denials all preserved
- TestCancellationAfterBudgetCheck (1): cancel set mid-call (after projection,
before commit) still honoured; output synthesised but state untouched
- TestCancellationAfterCommit (2): post-commit cancel not observable (honest
limit) BUT next call on same engine observes it + returns 'cancelled'
- TestLegacyCallersUnchanged (3): cancel_event=None preserves #162 atomicity
+ max_turns contract with zero behaviour change
- TestCancellationVsOtherStopReasons (2): cancel precedes max_turns check;
cancel does not retroactively override a completed turn
Runtime-layer (5 tests):
- TestTimeoutPropagatesCancelEvent (3): submit_message receives a real Event
object when deadline is set; None in legacy mode; timeout actually calls
event.set() so in-flight threads observe at their next checkpoint
- TestCancelEventSharedAcrossTurns (1): same event object passed to every
turn (object identity check) — late cancel on turn N-1 must affect turn N
Regression: 3 existing timeout test mocks updated to accept cancel_event
kwarg (mocks that previously had signature (prompt, commands, tools, denials)
now have (prompt, commands, tools, denials, cancel_event=None) since runtime
passes cancel_event positionally on the timeout path).
Full suite: 97 → 114 passing, zero regression.
Closes ROADMAP #164 Stage A.
What's explicitly NOT in Stage A:
- Preemptive cancellation of wedged provider IO (requires asyncio-native
provider path; larger refactor)
- Timeout on the legacy unbounded run_turn_loop path (by design: legacy
callers opt out of cancellation entirely)
- CLI exposure of 'cancelled' as a distinct exit code (currently 'cancelled'
maps to the same stop_reason != 'completed' break condition as others;
CLI surface for cancel is a separate pinpoint if warranted)
#159: multi-turn sessions had a silent security asymmetry: denied_tools
were always empty in run_turn_loop, even though bootstrap_session inferred
them from the routed matches. Result: any tool gated as 'destructive'
(bash-family commands, rm, etc) would silently appear unblocked across all
turns in multi-turn mode, giving a false 'clean' permission picture to any
claw consuming TurnResult.permission_denials.
Fix: compute denied_tools once at loop start via _infer_permission_denials,
then pass the same denials to every submit_message call (both timeout and
legacy unbounded paths). This mirrors the existing bootstrap_session pattern.
Acceptance: run_turn_loop('run bash ls').permission_denials now matches
what bootstrap_session returns — both infer the same denials from the
routed matches. Multi-turn security posture is symmetric.
Tests (tests/test_run_turn_loop_permissions.py, 2 tests):
- test_turn_loop_surfaces_permission_denials_like_bootstrap: Symmetry
check confirming both paths infer identical denials for destructive tools
- test_turn_loop_with_continuation_preserves_denials: Denials inferred at
loop start are passed consistently to all turns; captured via mock and
verified non-empty
Full suite: 82/82 passing, zero regression.
Closes ROADMAP #159.
The #160 session-lifecycle CLI triplet was asymmetric: list-sessions and
delete-session accepted --directory + --output-format and emitted typed
JSON error envelopes, but load-session had neither flag and dumped a raw
Python traceback (including the SessionNotFoundError class name) on a
missing session.
Three concrete impacts this fix closes:
1. Alternate session-store locations (e.g. /tmp/claw-run-XXX/.port_sessions)
were unreachable via load-session; claws had to chdir or monkeypatch
DEFAULT_SESSION_DIR to work around it.
2. Not-found emitted a multi-line Python stack, not a parseable envelope.
Claws deciding retry/escalate/give-up had only exit code 1 to work with.
3. The traceback leaked 'src.session_store.SessionNotFoundError' verbatim,
coupling version-pinned claws to our internal exception class name.
Now all three triplet commands accept the same flag pair and emit the
same JSON error shape:
Success (json mode):
{"session_id": "alpha", "loaded": true, "messages_count": 3,
"input_tokens": 42, "output_tokens": 99}
Not-found:
{"session_id": "missing", "loaded": false,
"error": {"kind": "session_not_found",
"message": "session 'missing' not found in /path",
"directory": "/path", "retryable": false}}
Corrupted file:
{"session_id": "broken", "loaded": false,
"error": {"kind": "session_load_failed",
"message": "...", "directory": "/path",
"retryable": true}}
Exit code contract:
- 0 on successful load
- 1 on not-found (preserves existing $?)
- 1 on OSError/JSONDecodeError (distinct 'kind' in JSON)
Backward compat: legacy 'claw load-session ID' text output unchanged
byte-for-byte. Only new behaviour is the flags and structured error path.
Tests (tests/test_load_session_cli.py, 13 tests):
- TestDirectoryFlagParity (2): --directory works + fallback to CWD/.port_sessions
- TestOutputFormatFlagParity (2): json schema + text-mode backward compat
- TestNotFoundTypedError (2): JSON envelope on not-found; no traceback in
either mode; no internal class name leak
- TestLoadFailedDistinctFromNotFound (1): corrupted file = session_load_failed
with retryable=true, distinct from session_not_found
- TestTripletParityConsistency (6): parametrised over [list, delete, load] *
[--directory, --output-format] — explicit parity guard for future regressions
Full suite: 80/80 passing, zero regression.
Discovered via Jobdori dogfood sweep 2026-04-22 17:44 KST — ran
'claw load-session nonexistent' expecting a clean error, got a Python
traceback. Filed #165 + fixed in same commit.
Closes ROADMAP #165.
#163: run_turn_loop no longer injects f'{prompt} [turn N]' into follow-up
prompts. The suffix was never defined or interpreted anywhere — not by the
engine, not by the system prompt, not by any LLM. It looked like a real
user-typed annotation in the transcript and made replay/analysis fragile.
New behaviour:
- turn 0 submits the original prompt (unchanged)
- turn > 0 submits caller-supplied continuation_prompt if provided, else
the loop stops cleanly — no fabricated user turn
- added continuation_prompt: str | None = None parameter to run_turn_loop
- added --continuation-prompt CLI flag for claws scripting multi-turn loops
- zero '[turn' strings ever appear in mutable_messages or stdout now
Behaviour change for existing callers:
- Before: run_turn_loop(prompt, max_turns=3) submitted 3 turns
('prompt', 'prompt [turn 2]', 'prompt [turn 3]')
- After: run_turn_loop(prompt, max_turns=3) submits 1 turn ('prompt')
- To preserve old multi-turn behaviour, pass continuation_prompt='Continue.'
or any structured follow-up text
One existing timeout test (test_budget_is_cumulative_across_turns) updated
to pass continuation_prompt so the cumulative-budget contract is actually
exercised across turns instead of trivially satisfied by a one-turn loop.
#164 filed: addresses reviewer feedback on #161. The wall-clock timeout
bounds the caller-facing wait, but the underlying submit_message worker
thread keeps running and can mutate engine state after the timeout
TurnResult is returned. A cooperative cancel_event pattern is sketched in
the pinpoint; real asyncio.Task.cancel() support will come once provider
IO is async-native (larger refactor).
Tests (tests/test_run_turn_loop_continuation.py, 8 tests):
- TestNoTurnSuffixInjection (2): zero '[turn' strings in any submitted
prompt, both default and explicit-continuation paths
- TestContinuationDefaultStopsAfterTurnZero (2): default loops run exactly
one turn; engine.submit_message called exactly once despite max_turns=10
- TestExplicitContinuationBehaviour (2): turn 0 = original, turn N = continuation
verbatim; max_turns still respected
- TestCLIContinuationFlag (2): CLI default emits only '## Turn 1';
--continuation-prompt wires through to multi-turn behaviour
Full suite: 67/67 passing.
Closes ROADMAP #163. Files #164.
Previously, QueryEnginePort.submit_message() checked the token budget AFTER
appending the prompt to mutable_messages, transcript_store, and permission_denials,
and AFTER calling compact_messages_if_needed(). On overflow it set
stop_reason='max_budget_reached' but the overflow turn was already committed.
Any caller that persisted the session afterwards wrote the rejected prompt to
disk — the session was silently poisoned even though the TurnResult said the
turn never completed.
Fix:
- Restructure submit_message so the budget check early-returns BEFORE any
mutation of mutable_messages, transcript_store, permission_denials, or
total_usage.
- The returned TurnResult.usage reflects pre-call state (overflow never
advanced the usage counter).
- Normal (in-budget) path unchanged: mutation happens exactly once, at the
end, only on 'completed' results.
This closes the atomicity gap: submit_message is now either 'turn committed'
(stop_reason='completed') or 'turn rejected, state untouched'
(stop_reason in {'max_budget_reached', 'max_turns_reached'}). Callers can
safely retry with a fresh budget or a smaller prompt without worrying about
phantom committed turns from prior rejections.
Tests (tests/test_submit_message_budget.py, 10 tests):
- TestBudgetOverflowDoesNotMutate (5): mutable_messages / transcript /
permission_denials / total_usage / TurnResult.usage all pre-mutation after overflow
- TestOverflowPersistence (2): first-turn overflow persists empty session;
successful-turn-then-overflow persists only the successful turn
- TestEngineUsableAfterOverflow (2): subsequent in-budget call still works
with no residue; repeated overflows don't accumulate hidden state
- TestNormalPathStillCommits (1): regression guard — non-overflow path still
commits mutable_messages/transcript/usage as expected
Full suite: 59/59 passing, zero regression.
Blocker: none. Closes ROADMAP #162.
Previously, run_turn_loop was bounded only by max_turns (turn count). If
engine.submit_message stalled — slow provider, hung network, infinite
stream — the loop blocked indefinitely with no cancellation path. Claws
calling run_turn_loop in CI or orchestration had no reliable way to
enforce a deadline; the loop would hang until OS kill or human intervention.
Fix:
- Add timeout_seconds parameter to run_turn_loop (default None = legacy unbounded).
- When set, each submit_message call runs inside a ThreadPoolExecutor and is
bounded by the remaining wall-clock budget (total across all turns, not per-turn).
- On timeout, synthesize a TurnResult with stop_reason='timeout' carrying the
turn's prompt and routed matches so transcripts preserve orchestration context.
- Exhausted/negative budget short-circuits before calling submit_message.
- Legacy path (timeout_seconds=None) bypasses the executor entirely — zero
overhead for callers that don't opt in.
CLI:
- Added --timeout-seconds flag to 'turn-loop' command.
- Exit code 2 when the loop terminated on timeout (vs 0 for completed),
so shell scripts can distinguish 'done' from 'budget exhausted'.
Tests (tests/test_run_turn_loop_timeout.py, 6 tests):
- Legacy unbounded path unchanged (timeout_seconds=None never emits 'timeout')
- Hung submit_message aborted within budget (0.3s budget, 5s mock hang → exit <1.5s)
- Budget is cumulative across turns (0.6s budget, 0.4s per turn, not per-turn)
- timeout_seconds=0 short-circuits first turn without calling submit_message
- Negative timeout treated as exhausted (guard against caller bugs)
- Timeout TurnResult carries correct prompt, matches, UsageSummary shape
Full suite: 49/49 passing, zero regression.
Blocker: none. Closes ROADMAP #161.
The old tracked TypeScript snapshot has been removed from the repository history and the root directory is now a Python porting workspace. README and tests now describe and verify the Python-first layout instead of treating the exposed snapshot as the active source tree.
A local archive can still exist outside Git, but the tracked repository now presents only the Python porting surface, related essay context, and OmX workflow artifacts.
Constraint: Tracked history should collapse to a single commit while excluding the archived snapshot from Git
Rejected: Keep the exposed TypeScript tree in tracked history under an archive path | user explicitly wanted only the Python porting repo state in Git
Confidence: medium
Scope-risk: broad
Reversibility: messy
Directive: Keep future tracked additions focused on the Python port itself; do not reintroduce the exposed snapshot into Git history
Tested: python3 -m unittest discover -s tests -v; python3 -m src.main summary; git diff --check
Not-tested: Behavioral parity with the original TypeScript system beyond the current Python workspace surface
YeonGyu-Kim
instructkr