Launch a `codeplain` render and supervise it end-to-end. Tails the log file, watches generated code appear under `plain_modules/`, and inspects the live TUI/process. Detects pathologies (stuck conformance loops, complexity errors, missing concepts, render failures), and — with user approval — stops the renderer, hands off to a spec-edit skill, and resumes with `--render-from`. Use whenever the user wants to actually render specs (not just dry-run) and wants supervision of the run.
Launch a `codeplain` render and supervise it end-to-end. Tails the log file, watches generated code appear under `plain_modules/`, and inspects the live TUI/process. Detects pathologies (stuck conformance loops, complexity errors, missing concepts, render failures), and — with user approval — stops the renderer, hands off to a spec-edit skill, and resumes with `--render-from`. Use whenever the user wants to actually render specs (not just dry-run) and wants supervision of the run.
Run Codeplain
Always use the skill load-codeplain-reference to retrieve the full codeplain CLI reference (every flag, path-resolution rules, config.yaml mapping, render banners, exit codes), and the skill load-plain-reference to retrieve the ***plain syntax rules — but only if you haven't done so yet. This skill drives the codeplain CLI, so the CLI reference is required before launching or resuming any render.
What this skill is
A supervisor around a live codeplain render. The renderer itself is the codeplain CLI; this skill does not generate code. It:
Launches codeplain <module>.plain (or attaches to an in-flight run).
Watches three signal sources in a tight loop, in priority order:
codeplain.log — the primary signal. Everything the renderer is doing, deciding, retrying, or failing at gets written here. This is the file you stare at.
the generated code outputs under plain_modules/<module>/ and conformance_tests/<module>/ — a secondary signal, used to corroborate or contradict what the log says.
the process / TUI itself (alive? exited? exit code?) — a tertiary signal, used only to know whether to keep monitoring.
Surfaces what is happening to the user in plain English, in near-real-time.
On pathology, stops the renderer (with user approval), hands off to the right edit skill, and resumes from the last completed functionality.
This skill is read-mostly. It does not modify .plain files itself — it delegates to debug-specs, resolve-spec-conflict, break-down-func-spec, add-implementation-requirement, etc. The only files it writes are control files: it may kill a process, and it shells out to codeplain again to resume.
The log is the ground truth. The TUI is opaque to the agent; the filesystem moves slowly and only on completed steps; but codeplain.log updates the moment the renderer takes any action — API call, retry, fix attempt, test run, abort. If the log and the filesystem disagree, trust the log. If the log and the TUI disagree (per the user's description), trust the log.
When to use
The user says "render it", "let's run codeplain", "kick off the render", "render <module>.plain", or similar.
The user already has a render going in another terminal and wants you to babysit it.
After forge-plain / add-feature / debug-specs / plain-healthcheck finished and the next move is the real render.
Do not use this skill for:
A dry-run gate — that is plain-healthcheck.
Authoring or editing specs — that is forge-plain, add-feature, debug-specs, or the specific add-* / resolve-* / break-down-* skills.
Pre-flight (always do this first)
Before launching anything:
Confirm the target. Ask the user which top module they want rendered if it is not unambiguous from the conversation. The argument is the path to a .plain file (e.g. agent_harness.plain, backend/api.plain).
Run plain-healthcheck if it hasn't been run since the last spec edit. A real render that fails on something a dry-run would have caught burns credits. If healthcheck FAILs, stop here and surface the failures — do not start the render.
Verify the environment.
codeplain is on PATH (command -v codeplain).
CODEPLAIN_API_KEY is exported, or the user will pass --api-key.
The governing config.yaml for the chosen module exists (use the pairing rule from plain-healthcheck: same-directory config wins; root config.yaml is the fallback).
Decide on flags with the user. Default to a minimal command line; mention these knobs only if they apply:
--config-name <name> — multi-part project, or non-default config file name.
--render-range <range> / --render-from <id> — partial render, or resume from a previous run.
--copy-conformance-tests / --conformance-tests-dest — same for conformance tests.
--test-script-timeout <seconds> — if test scripts are slow.
--api-key <key> / --api <url> — only if the env var route is not in use.
--render-machine-graph — only on explicit request.
--headless — see "Launch mode" below.
Do not prompt for flags the user did not signal interest in; the empty command line is the right default.
Locate and reset the log file. By default it is codeplain.log in the same directory as the .plain file. If --log-file-name was overridden, the override resolves based on where it was written: a relative value on the CLI resolves against the invocation directory; a relative value in config.yaml resolves against the config file's directory; absolute and ~ paths are used as-is. The file is overwritten at the start of each run (per the --log-file-name help text). So:
Before launch, note the absolute path of the log file and its current size (wc -c <path>) — this lets you detect the moment the new run truncates and starts writing.
After launch, every log read must be against this same path. Re-derive it (using the resolution rule above) if the user passes --log-file-name, sets log-file-name in config.yaml, or runs from a different working directory.
The log is per-run: every byte you read during this run was written by this run. You don't have to worry about stale lines from prior runs.
Read every test script referenced by the governing config.yaml. This is non-negotiable. The scripts are the contract between the renderer and the project; you cannot judge what the renderer is "fixing" without knowing what pass and fail actually mean in this project. For each of unittests-script, prepare-environment-script, and conformance-tests-script (those that are declared), read the script end-to-end before launching, and write down for yourself:
Which framework is invoked (e.g. vitest, jest, pytest, go test, cargo test, phpunit, etc.) and any framework flags (--noEmit, --reporter=verbose, custom configs).
What "pass" means, exit-code-wise. Some scripts combine multiple checks (e.g. unit_testing_typescript.sh here runs tsc --noEmitandvitest, and its final exit code is the worst of the two — so a type error alone fails the unit-test gate).
What "fail" means, including any failures that the script deliberately downgrades to warnings (e.g. prepare_environment_typescript.sh here treats a failed npm run build as a non-fatal warning). Failures the script swallows are failures the renderer will never see, so a spec defect that depends on them will silently slip through.
Where the scripts materialize their working environment (current-convention scripts stage into the system temp directory, e.g. /tmp/typescript_<build_folder_basename>/; older scripts used project-local folders like .tmp/typescript_<build_folder>/, build/, target/). This is where the renderer's iteration actually runs; if you need to look at what is breaking, this is the directory — not plain_modules/<module>/ directly.
How conformance tests are discovered and staged (glob, naming convention, alias setup, etc.). Some scripts stage tests into the source tree (this repo flattens conformance_tests/<module>/<fname>/ into src/<module>/); some run them in place; some require a specific file extension. This tells you exactly which conformance test file is in play for a given functionality.
Toolchain prerequisites the script asserts (Node version, Python version, specific binaries). If any are missing on the user's machine, stop now and tell them — the renderer will burn credits on phantom failures otherwise.
Keep this synthesis short (5–10 bullets, project-specific). It is the single most important reference for the classifier in Spec-deviation classification. The healthcheck only checks that the scripts exist and are referenced correctly; this step is the only place that reads what they actually do.
Launch always in --headless mode
The agent must always launch the renderer in the background and supervise it via the log file and filesystem. The terminal tool cannot drive an interactive TUI, so TUI mode is incompatible with agent-launched rendering. Always use --headless:
Record the PID. Confirm codeplain.log starts growing within a few seconds; if it doesn't, the launch failed — check the user's shell / API key and re-try.
Log monitoring contract
The monitor loop in Phase 1 leans almost entirely on codeplain.log. Before describing the loop, here is exactly how to read this file efficiently and what to look for. Treat this section as the reference the loop calls into.
How to read the log incrementally
The log grows monotonically during a run — only ever appended to, never rewritten — and can reach thousands of lines on a long render. Never re-read it whole on each pass. Track the last byte offset you read and read only the new bytes since then.
Keep two pieces of state across iterations:
LOG_PATH — absolute path to codeplain.log (captured during pre-flight).
LOG_OFFSET — number of bytes you have already consumed. Starts at 0 for a fresh run.
Each monitor pass:
SIZE=$(wc -c < "$LOG_PATH")
if [ "$SIZE" -gt "$LOG_OFFSET" ]; thentail -c +$((LOG_OFFSET + 1)) "$LOG_PATH"# new bytes only
LOG_OFFSET=$SIZEfi
If SIZE < LOG_OFFSET, the file was truncated — a new run has started or somebody rotated the log. Reset LOG_OFFSET=0 and re-read from the top.
If you cannot keep shell state between tool calls, the equivalent is: wc -c the file each pass to get the new size, remember the previous size in your own conversation state, and pass the byte range to tail -c +<offset+1> or dd skip=<offset> bs=1. Either way, never read the whole file on a routine pass — only when you need backstory (e.g. you just detected an error and want context from earlier).
Catalog of log patterns to recognize
All patterns below are taken from the actual codeplain runtime. Treat them as the canonical signal vocabulary.
Pattern (regex / substring)
Meaning
Loop action
^INFO:codeplain:[-]+$ followed by Module: <name> and Rendering functionality <N>:
A new functionality is starting. The block ends at the next ----- line.
Update CURRENT_MODULE and CURRENT_FUNCTIONALITY. Reset ATTEMPT_COUNTER to 0. Report to the user.
Running unit tests script .* \(attempt: <N>\)
Unit tests starting, possibly a retry.
If <N> climbs past 1, flag as warning; past 3, treat as Pathology G.
\[#79FC96\]All Unit Tests scripts have passed successfully.\[/#79FC96\]
Unit tests green.
Routine. Note the functionality is past the unit-test gate.
Refactoring the generated code...
Refactor pass between unit tests and conformance.
Routine.
Implementing conformance tests... / Implementing test requirements:
Conformance phase starting for the current functionality.
Routine.
Running testing environment preparation script .* for build folder
prepare-environment-script is running.
Routine.
\[#79FC96\]All Testing Environment Preparation scripts have passed successfully.\[/#79FC96\]
Env prep green.
Routine.
Running conformance tests script .* for conformance_tests/<m>/<fname> \(functionality <N> in module <m>\)
Conformance tests starting for functionality <N>.
Note <fname> — that is the on-disk folder where the latest conformance test lives.
Running conformance tests attempt <N>.
Conformance retry loop. Watch <N> carefully.
Increment ATTEMPT_COUNTER = <N>. At >=5, trigger Pathology A.
Fixing conformance test for functionality <N> in module <m>.
Renderer is patching its own conformance test between attempts.
On its own, routine. Combined with climbing <N>, smoking gun for Pathology A.
Functional spec too complex!
Single spec implies >200 LOC. Renderer aborts.
Pathology B.
Substrings: missing concept, unknown definition, cyclic, not defined, cannot resolve
Spec graph error. Renderer aborts.
Pathology C.
Substrings: conflict, conflicting, contradicts
Spec conflict surfaced.
Pathology D.
Traceback (most recent call last): followed by a Python stack
Renderer itself crashed (not a spec error).
Stop monitoring; surface the stack to the user verbatim; this is a bug report case, not a spec edit.
\[#FF6B6B\]✗ rendering failed
Renderer aborted. Followed by render id, input file, generated code folder, functionalities, used credits, render time.
Capture the whole failure block. Run is over — go to Phase 4 (failure branch).
\[#79FC96\]✓ rendering succeeded (or rendering complete) followed by the same metadata block
Renderer finished successfully.
Capture the success block. Run is over — go to Phase 4 (success branch).
Stop monitoring; tell the user the API rejected the request and surface the line.
No new bytes for > 2 min while the process is alive
API stall or wedged renderer.
Pathology E.
Where a column says "Pathology X", see Phase 2 for the response.
Three pieces of state to maintain across the loop
The pathologies in Phase 2 can only be detected if you carry state between iterations. Keep these three variables in your working memory across passes:
CURRENT_FUNCTIONALITY — the integer ID of the functionality currently being rendered. Updated whenever you see Rendering functionality <N>:.
LAST_COMPLETED_FUNCTIONALITY — the highest functionality ID whose conformance tests passed cleanly (i.e. no Fixing conformance test immediately after, and either the next Rendering functionality <N+1>: line appeared or the run succeeded). This is the value you'll pass to --render-from <N+1> on resume.
ATTEMPT_COUNTER — the most recent Running conformance tests attempt <N> value for CURRENT_FUNCTIONALITY. Reset to 0 when CURRENT_FUNCTIONALITY changes.
A single-line summary of these three values is what you'll show the user on each pass.
Useful one-liners
Use these as cheap, idempotent ways to interrogate the log. Each is safe to run on any pass.
grep -E "^(ERROR|CRITICAL|WARNING):codeplain:|✗ rendering failed|Functional spec too complex|Traceback" codeplain.log
Time the last log line was written (proxy for liveness):
stat -f %m codeplain.log # macOS — epoch mtime
Reading test-script output
The renderer drives three external shell scripts during a run — unittests-script, prepare-environment-script, and conformance-tests-script (whichever are declared in the project's config.yaml). Each script wraps a real test framework (vitest, jest, pytest, tsc, etc.) and writes verbose output. It reads the output those scripts produce while the renderer is iterating. Three places to look, in priority order:
1. Inline in codeplain.log (the primary source)
With verbose logging on (the default — --verbose is disabled), codeplain captures each script's stdout/stderr and writes it into codeplain.log between its own wrapper lines:
INFO:codeplain:Running unit tests script .../unit_testing_typescript.sh. (attempt: 1)
... script output here: vitest summaries, FAIL lines, tsc diagnostics ...
INFO:codeplain:[#79FC96]All Unit Tests scripts have passed successfully.[/#79FC96]
So when you do the incremental log read in Phase 1, you are simultaneously reading the renderer wrapper lines ("running", "passed", "fixing") and the raw framework output sandwiched between them. The wrapper says whether the script passed; the framework output says why it failed and what the renderer is reacting to. Both are needed.
What to extract from the framework output — stay framework-agnostic, look for these shapes:
Counts: <N> passed, <M> failed, <K> skipped. Climbing failure count between attempts means the renderer is making things worse.
Individual test names that failed: lines beginning with ×, FAIL, ✗, × should ..., FAIL src/.../foo.test.ts > <name>. The same test failing across attempts is the smoking gun for an under-specified spec.
Assertion bodies: expected X, got Y. The values reveal exactly what the test expects vs what the implementation produced — the cleanest input to the classifier below.
Compile / type errors (when the script wraps a type-checker): error TS<NNNN>, cannot find module, is not assignable to. A unit-test gate that keeps re-failing on the same type error means the implementation req or the type-level spec is wrong, not the code.
Stack traces: which file under the prepared working folder threw, and at which line. Map the file back to plain_modules/<module>/src/... to find the code the spec is supposed to be governing.
"No test files found" / "No tests collected": the script ran but discovered nothing. Almost always a path/glob mismatch in the staged working folder, not a spec problem. Treat as Pathology F.
Always set the logs to be verbose in config.yaml.
2. Renderer memory under plain_modules/<module>/.memory/
The renderer writes structured JSON notes about its own iteration there. The most useful is conformance_test_memory/<issue_id>.json, with fields like:
{"functionality":"Agent CLI","initial_issue_summary":"...what was failing at the start...","fix_attempt_summary":"...what the renderer changed...","current_issue_summary":"...what is still wrong now...","resolution_status":"UNRESOLVED","key_learnings":"...the renderer's own diagnosis of why..."}
This is the renderer telling you, in its own words, what it tried and what is stuck. A resolution_status: "UNRESOLVED" is a hard signal that the loop is not converging — read the key_learnings field; it usually pinpoints the ambiguity in the spec or the limit of what the renderer can infer.
Refresh these files on each pass while a retry loop is active — the renderer rewrites them between attempts.
Spec-deviation classification
This is the most important active judgment the skill makes. Once a functionality enters a retry loop (unit-test or conformance), the renderer iterates — changing source code, and sometimes changing its own conformance tests — trying to satisfy the test-script output. Not every iteration is honest. Each iteration must be classified before deciding whether to let the renderer continue, stop and fix the spec, or stop and fix the test script.
The inputs to the classifier are:
The spec that governs the failing functionality — the functional spec and its acceptance test(s) in the .plain file.
The test-script output for the most recent attempt, from codeplain.log per the section above.
The renderer's own .memory/conformance_test_memory/<id>.json for this functionality (when present).
The newest conformance test file the renderer just edited under conformance_tests/<module>/<fname>/.
The newest implementation file(s) the renderer just edited under plain_modules/<module>/.
With these in hand, place the current iteration into one of four buckets:
Bucket 1 — Honest convergence (let it continue)
Signals:
Failure count is going down across attempts (e.g. 5 failed → 3 failed → 1 failed).
The set of failing test names is shrinking and is a subset of the previous attempt's failures.
The conformance test file is unchanged between attempts — only the implementation is changing.
key_learnings in memory (if present) reads as legitimate technical reasoning about the framework / language / library, not about reinterpreting the spec.
Action: let the renderer keep going, up to the threshold (5 conformance attempts / 3 unit-test attempts). Report progress to the user with the per-attempt failure count.
Bucket 2 — Spec under-specification (stop and fix the spec)
Signals — any one of these is enough:
The same test fails every attempt with the same assertion, and the assertion checks something the functional spec does not explicitly say (e.g. exact exit code, exact error wording, specific status field).
current_issue_summary in memory describes a question the spec does not answer (e.g. "unclear whether --help should exit 0 or 2").
The renderer's fix_attempt_summary keeps oscillating between two opposite interpretations of the same requirement (regex variant A this attempt, variant B next attempt, A again the attempt after).
The conformance test file is being rewritten between attempts in a way that changes which behavior is being asserted, not just how it is asserted.
Action: stop the renderer (SIGINT). Hand off to debug-specs. The spec must be tightened so there is only one valid behavior to converge on. Then resume with --render-from <N> for the offending functionality.
Bucket 3 — Renderer is drifting away from the spec (stop immediately)
The dangerous case. Signals:
The renderer rewrote a conformance test to weaken the assertion — the test now accepts behavior the original spec does not allow. Compare the latest conformance test file to the previous attempt's; an assertion that loosened a === to a >=, removed an expect, or wrapped behavior in a try/catch to swallow failures is a drift signal.
The renderer deleted code that implemented a requirement to make a test pass (a try/catch that now silently returns 0, a feature flag short-circuit, a hard-coded return value matching the test). Visible by diffing plain_modules/<module>/src/... across attempts.
The renderer's key_learnings reads like rationalization rather than diagnosis (e.g. "adjusted assertion to match observed behavior", "removed strict check that was causing failures").
Failure count is going up between attempts.
Action: stop the renderer immediately, don't wait for the threshold. The fix must be applied to the .plain file (the spec was probably not specific enough and the renderer chose the path of least resistance):
If it happened during the conformance testing phase, the ***test reqs*** need to be tightened.
If it happened during the rendering of a functional requirement, the functional specs, definitions, or ***impl reqs*** need to change.
If it happened during acceptance testing, the acceptance test definitions need to be tightened.
Do nothing on the generated-code side; the broken code under plain_modules/ will be overwritten on resume. Never hand-edit it.
Then resume with --render-from <N> (or --force-render if the drift contaminated earlier functionalities). Re-running plain-healthcheck first is mandatory.
Bucket 4 — The test script itself is the source of the failure (Pathology F)
Signals:
Test-script output shows a toolchain error (vitest: command not found, python: No module named pytest, Node version too old).
Output is No test files found / no tests collected while the conformance tests do exist on disk — a glob/path/staging bug in the script.
The script is failing before running any framework command at all (die: prepared environment missing, missing argument, etc.).
The failure mode is identical across functionalities, not specific to the spec being rendered.
Action: This is not a spec problem and the spec should stay as is. Stop the renderer and investigate why the script fails:
If it's a bug in the script (e.g., pathing issue, syntax error, missing environment variable load), attempt to fix the script directly, or hand off to the matching implement-*-testing-script skill.
If it's a missing host dependency (e.g., Python or Node is not installed, or a required global package is missing), inform the user and ask them to install it.
Before resuming the renderer, test the script by running it manually exactly as the renderer would. If the same infrastructure/script issue persists, the fix didn't work. If the script now runs (even if it reports legitimate test failures), the fix was successful.
Classifier discipline
Run the classifier only when a retry loop is actually in progress (ATTEMPT_COUNTER >= 2). For a first attempt that just failed, there isn't enough data yet — the renderer hasn't done a fix pass.
Show your classification reasoning to the user before acting on Bucket 2 or Bucket 3. State the bucket, the concrete evidence ("test should exit 0 on --help has failed 3 attempts; spec text does not specify exit code"), and the proposed hand-off skill. Get their go-ahead before SIGINT.
If two buckets seem to apply, prefer Bucket 4 → Bucket 3 → Bucket 2 → Bucket 1 in that order. A script bug shadows everything; a drift is worse than under-specification; under-specification is worse than honest iteration.
Phase 1 — Monitor loop
Run the loop until one of these terminal conditions holds: the process exits successfully, the process exits with failure, or you (with the user's approval) kill it.
On each iteration, in this order:
1a. Read the new log bytes (primary)
Using the incremental read recipe above, pull only the bytes appended since the last pass. Walk the new text against the pattern catalog and update CURRENT_FUNCTIONALITY, LAST_COMPLETED_FUNCTIONALITY, and ATTEMPT_COUNTER. Because verbose mode is on by default, this chunk also contains the raw test-script output (vitest summaries, tsc errors, stack traces) sandwiched between the renderer's wrapper lines — capture failure counts, failing test names, and key assertion text. Note any pathology hits, but don't act on them yet; finish the pass first so the report is consistent.
If there are no new bytes, note the elapsed time since the last log activity (stat mtime works). Hitting 2 minutes of silence with the process alive triggers Pathology E.
1b. Is the process alive?
Mode A: ps -p <pid> (or kill -0 <pid>). If gone, capture the exit code via wait, or fall back to the log's final block (success / failure banner).
Mode B: pgrep -fl codeplain should still return the run. If not, the user already stopped it (or it finished).
If the process is gone but you haven't yet seen a ✓ or ✗ banner in the log, do one more incremental read — the banner is the very last thing codeplain writes, and the process exits immediately after.
1c. Inspect the filesystem (corroboration only)
This is a secondary signal — it confirms what the log already said. Don't go on a fishing trip here.
For the current module, the renderer drops generated code into plain_modules/<module>/ and conformance tests into conformance_tests/<module>/<functionality_name>/. Each pass, inspect what is new:
New files under plain_modules/<module>/src/ confirm the functionality is being implemented.
New folders under conformance_tests/<module>/ confirm the functionality has reached the conformance phase.
Files that were rewritten between passes are the renderer's "fix" attempts — diffing them across iterations is what reveals whether the renderer is converging or thrashing.
Read only the newest file(s) (the ones that changed since last pass), and read them to understand what the renderer chose to do, not just that it did something. This is the first line of defense against "the spec is ambiguous and the renderer guessed wrong" — you can often spot the wrong guess in the generated code long before tests fail.
Generated code under plain_modules/ and conformance_tests/ is read-only. Never edit it. Edits go in the .plain files.
1d. Cadence
Poll every 10–30 seconds. Faster than 10s is noisy and wastes tool calls; slower than 30s makes the user feel ignored. Tighten the cadence (down to 5s) when something looks off — climbing ATTEMPT_COUNTER, a new error keyword in the latest log chunk, an unexpectedly quiet log. Loosen the cadence (out to 60s) during long, healthy "rendering functionality N" stretches where the log is moving steadily and nothing alarming has appeared.
The log read is cheap (incremental), so the bottleneck is the user-facing report, not the I/O. Prefer to err on the tighter side.
1e. Report to the user
After each pass, give the user a short status line — one or two sentences, built from the three state variables. Examples:
Functionality 3/? in flight; unit tests just passed, conformance prep starting. (after a Implementing conformance tests... line)
Functionality 4 is on conformance attempt 4 of an open loop; unit tests still green. One more attempt and I'll flag it. (ATTEMPT_COUNTER = 4)
No new log lines in 95s; PID 12345 still alive. Likely an API call in flight. (silent log, alive process)
Last completed: functionality 2. Currently rendering functionality 3. (resume-readiness summary)
Do not dump the raw log unless the user asks for it. When you do dump it, paste only the relevant block (e.g. the failure banner with the surrounding 10 lines), not the entire file.
Phase 2 — Pathologies and intervention
These are the patterns that justify stopping the renderer. For each, the workflow is the same: detect → confirm with the user → stop → diagnose → fix the specs → resume.
Pathology A — Conformance loop that won't converge
Symptom: Running conformance tests attempt <N> keeps climbing for the same functionality, with Fixing conformance test for functionality <N> between attempts. Unit tests stay green; only the conformance step fails.
Threshold: 5 attempts on the same functionality is the heuristic. The renderer will keep going (the real codeplain.log in this repo shows it climbing to 10), but each attempt costs credits and rarely converges if it hasn't by 5.
Do not treat hitting the threshold as the diagnosis. From ATTEMPT_COUNTER >= 2 onwards, run the Spec-deviation classification classifier each pass. The bucket it returns is the diagnosis:
Bucket 1 — keep going until the threshold; report the per-attempt failure count.
Bucket 2 — stop now (regardless of how close to the threshold you are); hand off to debug-specs / add-implementation-requirement / inline spec edit. The spec is under-specified.
Bucket 3 — stop immediately, do not wait for the threshold; the renderer is drifting. Tighten the spec or acceptance test before resuming.
Bucket 4 — the test script is the problem; route to Pathology E instead.
Pathology B — Functional spec too complex!
The renderer rejected a spec because it would imply >200 LOC. Stop is automatic — the renderer aborts. Hand off to break-down-func-spec.
The renderer can't resolve a reference. Stop is automatic. Hand off to add-concept (or fix the import / requires chain). Re-run plain-healthcheck before resuming — a dry-run will catch this faster next time.
Pathology D — Conflicting specs
Renderer surfaces a conflict, or you can see two specs in the same module that contradict each other in the generated code. Hand off to resolve-spec-conflict.
Pathology E — Test scripts themselves are broken
Symptom: log shows the unit-test or conformance-test script itself failing (non-zero exit, syntax error, missing dependency) on every attempt, rather than the renderer's code failing the tests. The fix is not a spec change — it is a test_scripts/ change. Stop and hand off to the matching implement-*-testing-script skill, or have the user repair the script directly.
Special case — stale scripts after a renderer upgrade (doubled paths). The renderer now passes the build folder and conformance-tests folder to the scripts as absolute paths; older test scripts built paths by concatenating those arguments (.tmp/$1, python_$1, $current_dir/$2). The tell is a test-log path that contains the project root twice, e.g.:
ImportError: Start directory is not importable
/Users/x/project//Users/x/project/conformance_tests/module/spec
The render then gets stuck "fixing" tests that can never pass. The fix is never a spec edit: stop the render and regenerate the affected scripts via the matching implement-*-testing-script skill (current convention: system-temp working folder from basename "$1", $2 resolved to absolute before any cd).
Pathology F — Unit-test retry loop
Less common than the conformance loop, but the same shape: Running unit tests script ... (attempt: <N>) with <N> climbing for the same functionality. Threshold: 3 attempts. From ATTEMPT_COUNTER >= 2 onwards, run the same Spec-deviation classification classifier and act on the bucket it returns. The most common root causes for unit-test loops are an implementation req that doesn't match the unit-test scaffolding (Bucket 2), or test reqs / framework usage that the renderer is satisfying by weakening assertions (Bucket 3). Hand-off is typically add-implementation-requirement or debug-specs.
How to actually stop the renderer
Always SIGINT first, never SIGKILL unless SIGINT fails:
kill -INT <pid>
SIGINT lets codeplain flush the log and close out the run cleanly. Wait ~5s, then check ps -p <pid>. If still alive, repeat once. Only after a second SIGINT has failed should you escalate to kill -TERM <pid> and, last resort, kill -9 <pid>.
After the process is gone, do one final incremental read of the log — the renderer often writes a closing block on shutdown (✗ rendering failed with render id, input file, used credits, render time). Capture it; it's the most useful single artifact for the user.
The value of LAST_COMPLETED_FUNCTIONALITY you've been tracking is what you'll feed --render-from on resume (as <N+1>). If you weren't tracking it carefully, recover it now with:
and subtract 1 if the last functionality didn't reach a clean conformance pass.
Phase 3 — Fix + resume
Hand off to the right edit skill. Don't do the spec edit inside this skill — invoke the dedicated one (debug-specs, resolve-spec-conflict, break-down-func-spec, add-implementation-requirement, add-concept, add-functional-spec, etc.). Pass it concrete evidence: the offending log lines, the file paths under plain_modules/ you read, and the spec the user agreed is wrong.
Re-run plain-healthcheck after the fix. If it FAILs, do not resume — fix the next thing first.
Resume with --render-from. If the last fully-completed functionality was <N>, resume with the next one:
--render-from includes its argument, so the value passed is the first functionality you want re-rendered, not the last one that already passed. If nothing completed, omit --render-from and start over from the beginning. If only a single functionality needs re-rendering and the rest are fine, prefer --render-range <N> instead.
Re-enter the monitor loop in Phase 1.
If the fix required regenerating a module's earlier output (e.g. a definition change that ripples backwards), use --force-render instead of --render-from. Confirm this trade-off with the user — it costs more credits.
Phase 4 — Wrap up
When the renderer exits successfully:
Read the success banner from the log: render id, generated code folder, functionalities count, used credits, render time. Report it to the user verbatim — these are the numbers they actually care about.
List the top-level files/folders that appeared (or changed) under plain_modules/<module>/ and conformance_tests/<module>/ so the user knows where to look.
Remind the user of the side-channel commands they may want to run themselves, based on the config:
./test_scripts/<unittests-script> for unit tests,
./test_scripts/<prepare-environment-script> to set up the test env,
./test_scripts/<conformance-tests-script> for conformance tests.
Only mention the scripts the project's config.yaml actually declares.
If --copy-build was set, confirm the build was copied to --build-dest. If --copy-conformance-tests was set, same for --conformance-tests-dest.
When the renderer exits with a failure that wasn't intercepted by Phase 2 (e.g. transient API error, auth failure), report the failure block from the log and ask the user whether to retry (--render-from from the last completed ID) or to investigate.
Anti-patterns
Editing generated code to "fix" a render. Never. Generated files under plain_modules/ and conformance_tests/ exist as evidence only; they are overwritten every render. All fixes go in .plain files.
Reading the test scripts' source code to understand what's failing. Don't. The scripts run real frameworks (vitest, pytest, etc.) and the output of those frameworks — captured in codeplain.log and in .memory/conformance_test_memory/*.json — is what tells you what is failing and why. Reading the script source tells you nothing the log doesn't already say.
Treating every conformance loop as under-specification. Run the classifier. The renderer sometimes drifts (Bucket 3) by silently weakening its own tests; that is the opposite problem and the fix is different.
Re-reading the whole log on every pass. It grows. Track the byte offset and read only what was appended since the last pass. Full reads are reserved for "I just saw an error and need backstory."
Trusting the filesystem over the log. The filesystem only updates when a step completes; the log updates the moment any step starts, retries, or fails. When they disagree, the log is right.
Letting a conformance loop run indefinitely. Each attempt costs credits and the renderer will keep going past any reasonable point (this repo's codeplain.log climbed to attempt 10 before bailing). The threshold is 5; surface it to the user immediately.
Re-running the whole module after a small fix. Use --render-from <N+1> or --render-range <N>. Only fall back to --force-render when a backward dependency genuinely changed.
Trying to read the TUI. You can't see it. Use the log file and filesystem; ask the user to paste the TUI's test panel if the log is missing verbose output.
Dumping raw log into chat on every pass. Summarize from the three state variables (CURRENT_FUNCTIONALITY, LAST_COMPLETED_FUNCTIONALITY, ATTEMPT_COUNTER) plus a classifier verdict when a loop is in progress. The user wants signal, not transcript.
Skipping plain-healthcheck because "dry-run passed earlier". Specs drift; configs drift. Re-run it before every real render and after every fix.
kill -9 as a first move. Always SIGINT first — codeplain needs to flush its log and clean up.
Validation checklist
Target .plain file and governing config.yaml were confirmed before launch
plain-healthcheck is PASS as of the latest spec state
CODEPLAIN_API_KEY is set, or --api-key was supplied
Launch mode (--headless agent-launched vs. user-launched TUI) was chosen with the user
Renderer PID is known (or, in Mode B, confirmed alive via pgrep)
Log file path is known and is observed to be growing
Log path was captured during pre-flight and a byte offset is being tracked across passes
Monitor loop reads only the new bytes of codeplain.log each pass (no full re-reads on routine passes)
CURRENT_FUNCTIONALITY, LAST_COMPLETED_FUNCTIONALITY, and ATTEMPT_COUNTER are maintained across passes
Filesystem inspection is used only to corroborate log signals, not as the primary signal
Status reports to the user are short (≤2 sentences) and signal-driven, not transcript dumps
Test-script output (failure counts, failing test names, assertion bodies, stack traces) is being extracted from the verbose log chunks each pass
When ATTEMPT_COUNTER >= 2, the spec-deviation classifier ran and returned a bucket (1–4) before any stop/continue decision
If a classifier verdict caused a stop, the bucket and concrete evidence (failing test name + spec gap, or the diff that showed drift) was shown to the user before SIGINT
.memory/conformance_test_memory/<id>.json was read (if present) to capture current_issue_summary and resolution_status
Pathologies (loop / complexity / missing concept / conflict / stall / broken test script / drift) are recognized and acted on
Renderer was stopped with SIGINT (escalating only on failure) — not SIGKILL by default
Spec fixes were delegated to the right edit skill, not done inside this skill
Resumes use --render-from <N+1> (or --render-range) — --force-render only with user approval
On success, final render id / credits / time were reported and side-channel test commands were surfaced
No generated code under plain_modules/ or conformance_tests/ was modified
