Anchor Repro Agent Skill

anchor-repro

Force the skipped move in debugging: observe the bug before editing, then prove

the fix with the same observation.

Trigger Check

Use this skill when the task is to correct observable behavior that is currently

wrong: a failing test, error message, stack trace, regression, crash, hang, wrong

return value, wrong render, or "X does Y; it should do Z."

Do not use it for new features, refactors, renames, formatting, docs/copy

changes, dependency bumps without a behavioral symptom, or performance work

without a measurable probe.

Boundary test: can the wrongness be written as "running X currently produces Y;

it should produce Z"? If yes, use this skill and make X the probe seed. If the

statement cannot be formed from current information, ask for or find the missing

observation before editing.

Forced Procedure

• Capture the symptom and stable anchor. Choose the most stable literal

substring of the observable symptom for --expect. Omit heap addresses,

PIDs, timestamps, temp paths, and wrapped fragments. Use --expect whenever

the symptom has stable text. If no stable text exists, --no-stable-expect

is the honest fallback, not an unanchored run; do not use it when stable

textual output exists.

• Locate, then probe. Read enough code to place the behavior. Pick the

smallest CLI probe that exercises the reported symptom. In workspaces or

monorepos, run from the package with --cwd <package-dir>, not with a

cd ... && convention. If required environment changes matter, encode them

inside --shell as one quoted command string such as

'VAR=... command'; inherited environment changes are not part of probe

identity in v1.

• Demonstrate. Run the probe through

scripts/repro_ledger.py run --project <root> --bug <slug> ... -- <command>.

Read the failing output and confirm it fails for the reported reason:

mechanically, expectMatched should be true when --expect was supplied;

procedurally, inspect the tail/context instead of trusting an exit code.

• Diagnose against the probe. Test hypotheses with discriminating runs

before stacking edits. If a better probe is needed, the new probe starts a

new fail -> pass cycle and must be observed failing too.

• Fix.
• Verify with the identical probe identity. Re-run the same execution mode,

argv vector or shell string, --cwd, and --timeout. Then run

scripts/repro_ledger.py close --project <root> --bug <slug>.

• Promote when possible. If the target repo has a test suite, promote the

probe into a committed regression test. This is encouraged, not a gate.

• Report. Cite the close JSON, cite the first repro entry's git evidence,

name the strength rung from reference.md, and carry speculative when the

ledger took an unreproduced path.

Ledger Usage

Run and record a probe:

python3 skills/anchor-repro/scripts/repro_ledger.py run \
  --project /path/to/project \
  --bug login-crash \
  --cwd packages/app \
  --expect "TypeError: cannot read" \
  --timeout 60 \
  -- npm test -- login.test.ts

Verify after the fix:

python3 skills/anchor-repro/scripts/repro_ledger.py run \
  --project /path/to/project \
  --bug login-crash \
  --cwd packages/app \
  --expect "TypeError: cannot read" \
  --timeout 60 \
  -- npm test -- login.test.ts

python3 skills/anchor-repro/scripts/repro_ledger.py close \
  --project /path/to/project \
  --bug login-crash

Mark an honestly speculative fix only when reproduction is impossible:

python3 skills/anchor-repro/scripts/repro_ledger.py mark-unreproduced \
  --project /path/to/project \
  --bug customer-only-crash \
  --no-probe-possible \
  --reason "Requires customer-only data and credentials unavailable in this environment"

The ledger writes target-project artifacts under .techne/repro/ and

idempotently adds .techne/ to the target project's .gitignore. Do not commit

.techne/ output.

Stop Conditions

• Stop before editing if you have not run a failing probe or recorded an

explicit mark-unreproduced path.

• Stop if a failing run does not match the reported symptom; a hollow failure is

not a reproduction.

• Stop if a passing post-fix run uses a different identity; changed probes need

their own fail -> pass cycle.

• Stop before claiming a non-speculative fix if close does not exit 0 with a

verified same-probe summary.