cleveragents-core/tools/implementer_pr_context.py

#!/usr/bin/env python3
"""Worker-side reader for the dispatcher's PR-context handoff sentinel.

What this is
------------

The dispatcher pre-fetches everything the implementer worker needs to
make a code change — PR description, full diff, CI status + per-check
detail, PR comments, active REQUEST_CHANGES reviews, linked issues,
parent Epic body — and writes the lot to a sentinel at
``/tmp/cleveragents-implementer-handoff/pr-{N}.json``. The dispatcher
ALSO embeds those sections in the wrapper session's user prompt, but
each ``task`` tool call inside OpenCode re-summarises the prompt for
the child agent — by the time the chain reaches ``task-implementor``
at depth 3, only the diff section reliably survives.

This script is what the implementer worker calls (via the
``implementer-pr-context`` skill) to re-read any prefetched section
on demand, regardless of how the prompt got mangled upstream. It is
the read half of :mod:`tools._pr_context_sentinel`.

Contract
--------

Subcommand ``read``:

  ``python3 tools/implementer_pr_context.py read --pr <N> --field <name>``

  Reads the sentinel at ``/tmp/cleveragents-implementer-handoff/pr-{N}.json``,
  pulls the requested field, prints it on stdout, and exits 0. The
  ``--field`` value selects which slice of the prefetched data:

  - ``description`` — the PR body text (plain string)
  - ``title`` — the PR title (plain string)
  - ``metadata`` — JSON with work_type / work_group / head_sha /
    head_ref / base_ref / pr_number / data_complete / error_kinds
  - ``ci`` — JSON with the latest CI ``state`` plus ``checks[]``
    (per-status detail, including ``target_url`` and any prefetched
    failure-log excerpts)
  - ``comments`` — JSON list of PR / issue comments (chronological).
    For PRs this is the BOUNDED view: the most-recent N comments plus
    every human/reviewer comment — the older bot ``**Implementation
    Attempt**`` comments are rolled up into ``comments_digest`` instead.
  - ``comments_digest`` — JSON dict: the deterministic attempt-history
    digest (counts by tier / outcome, gates failing most, last
    success) computed from the FULL PR comment list. ``{}`` when no
    attempt comments were parsed.
  - ``reviews`` — JSON list of active REQUEST_CHANGES reviews + their
    inline comments (for ``request_changes_pr`` work only)
  - ``issues`` — JSON list of linked issues (the ``Closes #N``
    targets, with their bodies inlined)
  - ``epic`` — JSON dict with the parent Epic body, or empty if no
    Epic reference was found
  - ``diff`` — the unified diff as plain text (between
    ``BEGIN_PR_DIFF`` / ``END_PR_DIFF`` markers in the original
    prompt)
  - ``compliance_gaps`` — JSON dict of the dispatcher's deterministic
    PR Compliance Checklist scan (worktree_clean, changelog,
    contributors, issues-closed footer). Empty stdout when the
    dispatcher did not run the scan this cycle.
  - ``gate_preflight`` — JSON dict of the heavy flaky-test gate
    pre-flight (two ``local_ci_gate.sh --fast`` runs, classification
    of persistent vs. flaky failures). Empty stdout when the
    dispatcher did not run the pre-flight this cycle.
  - ``all`` — the entire sentinel payload as JSON (useful for
    debugging; ``task-implementor`` should pull narrowly)

  The script emits one of three signals on stdout, all with exit
  code 0:

  - **Empty stdout** — the sentinel is missing, malformed, or the
    dispatcher's fetch for this field did not complete (per the
    matching ``*_completed`` flag). The worker should fall through
    to its legacy curl-based GET against the live Forgejo API.
  - ``null\\n`` — the dispatcher fetched successfully and
    authoritatively confirmed absence (e.g. the PR has no Epic
    reference). The worker should SKIP its legacy GET; there is
    nothing to fetch.
  - Any other content — the field's value (plain text for
    description/title/diff/issue_body, pretty-printed JSON for
    everything else). The worker should use it verbatim.

  This three-case contract is how the worker distinguishes "the
  dispatcher didn't try" (fall through) from "the dispatcher tried
  and confirmed there is nothing" (don't re-curl). Conflating the
  two — as the original ``or None`` projection did — caused the
  worker to redundantly hit Forgejo for sections the dispatcher
  had already verified empty.

Exit codes
----------

- 0: success (or empty stdout for "no data available" — see above).
- 2: usage error (bad / missing CLI arg, unknown ``--field``).
- 3: schema-version mismatch (a config issue the operator must
  notice; the worker should treat this as "no data available" but
  the non-zero code makes the situation visible in logs).

Allow-rule
----------

This script's path is whitelisted under
``python3 tools/implementer_pr_context.py *`` in the agent's
permission allowlist (see :file:`.opencode/agents/task-implementor.md`).
"""
from __future__ import annotations

import argparse
import json
import os
import sys
from pathlib import Path
from typing import Any


# Must match :data:`tools._pr_context_sentinel.SCHEMA_VERSION`.
EXPECTED_SCHEMA_VERSION = 1

_DEFAULT_HANDOFF_DIR = Path("/tmp/cleveragents-implementer-handoff")


def _handoff_dir() -> Path:
    """Resolve the on-disk directory. Mirrors
    :func:`tools._pr_context_sentinel.handoff_dir`."""
    return Path(
        os.environ.get("IMPLEMENTER_DISPATCHER_PR_CONTEXT_DIR")
        or str(_DEFAULT_HANDOFF_DIR)
    )


def _handoff_path(pr_number: int) -> Path:
    return _handoff_dir() / f"pr-{int(pr_number)}.json"


def _load(pr_number: int) -> tuple[dict[str, Any] | None, str | None]:
    """Read + parse + validate the sentinel.

    Returns ``(payload, error)``:

    - ``(payload, None)`` — sentinel exists, parses, schema matches.
    - ``(None, None)`` — sentinel does not exist (normal: dispatcher
      didn't prefetch, or already cleaned up).
    - ``(None, error)`` — sentinel exists but is unusable (malformed,
      schema mismatch). The caller emits the error to stderr.
    """
    target = _handoff_path(pr_number)
    if not target.exists():
        return None, None
    try:
        raw = target.read_text(encoding="utf-8")
        payload = json.loads(raw)
    except (OSError, ValueError) as e:
        return None, f"handoff read/parse failed at {target}: {e}"
    if not isinstance(payload, dict):
        return None, f"handoff at {target} is not a JSON object"
    schema = payload.get("schema_version")
    if schema != EXPECTED_SCHEMA_VERSION:
        return None, (
            f"handoff at {target} has schema_version={schema!r}, "
            f"expected {EXPECTED_SCHEMA_VERSION}"
        )
    return payload, None


# Sentinel for "field is not present in this handoff" / "fetch failed
# upstream". Distinct from ``None`` so the worker contract can
# distinguish three cases:
#
# - ``_MISSING`` → emit NOTHING (empty stdout); worker falls through
#   to its legacy GET. Used when the dispatcher didn't attempt this
#   section (legacy prompt path) OR when the dispatcher attempted
#   and failed (transient error reflected in the matching
#   ``*_completed=False`` flag).
# - ``None`` → emit the JSON literal ``null`` (stdout = ``null\n``).
#   Used when the dispatcher fetched successfully and confirmed
#   absence (e.g. the PR has no Epic reference). The worker sees
#   non-empty stdout and skips its legacy GET, treating "null" as
#   the authoritative "no data" answer.
# - any other value → emit it; worker uses it verbatim.
#
# Without this distinction the worker re-curls the live Forgejo API
# every time the dispatcher confirmed an absence, defeating the
# whole point of the handoff.
_MISSING: Any = object()


def _gated_list(
    payload: dict[str, Any], value_key: str, completed_key: str
) -> Any:
    """Three-case helper for list-valued fields. ``_MISSING`` when
    ``completed_key`` is False; the value (or ``[]`` if absent /
    falsy) otherwise. Empty list serialises as ``[]\\n``, which
    the worker reads as authoritative-empty (skip legacy GET)."""
    if not payload.get(completed_key, True):
        return _MISSING
    return payload.get(value_key) or []


def _gated_text(
    payload: dict[str, Any], value_key: str, completed_key: str
) -> Any:
    """Three-case helper for plain-text fields.

    - completion flag False → ``_MISSING`` (worker falls through to
      legacy GET).
    - completion flag True + empty body → ``None`` (rendered as
      ``null\\n``; worker reads as authoritative-empty and skips
      the legacy GET).
    - completion flag True + non-empty body → the body verbatim.

    The completion-flag-aware path is what lets a legitimately-empty
    PR description (``description_completed=True``, ``description=""``)
    cleanly signal "no description, don't re-fetch" rather than
    forcing a redundant GET. Before the flag existed, every empty
    description / title / diff / issue_body cost the worker a
    redundant Forgejo round-trip.
    """
    if not payload.get(completed_key, True):
        return _MISSING
    body = payload.get(value_key)
    if not body:
        return None
    return body


def _project_field(payload: dict[str, Any], field: str) -> Any:
    """Return the slice of ``payload`` selected by ``--field``.

    Implements the **three-case contract** the worker relies on:

    - :data:`_MISSING` — the field is genuinely absent OR the
      dispatcher's fetch for it did not complete successfully (per
      the field's ``*_completed`` flag). :func:`_emit` renders this
      as empty stdout; the worker falls through to its legacy GET.
    - ``None`` — the dispatcher's fetch succeeded AND confirmed
      the field is empty in a way the JSON schema can encode (e.g.
      ``epic=None`` means "PR has no parent Epic"). :func:`_emit`
      renders this as ``null\\n``; the worker skips the legacy GET.
    - any other value — the value itself, rendered verbatim (plain
      text for ``description``/``title``/``diff``/``issue_body``,
      pretty-printed JSON for the rest). Empty lists / empty dicts
      ALSO fall through this case and render as ``[]\\n`` / ``{}\\n``
      — they carry the "fetched, confirmed empty" semantics for
      list/dict-valued fields.

    The :data:`_MISSING` vs ``None`` vs value distinction is what
    stops the worker from re-curling Forgejo for sections the
    dispatcher has already confirmed are empty. Mapping all three
    cases to ``None`` (the pre-2026-05-10 behavior) collapsed the
    "didn't try" and "tried, confirmed empty" cases together and
    every confirmed-empty section cost the worker a redundant GET.
    """
    if field == "all":
        return payload
    if field == "description":
        return _gated_text(payload, "description", "description_completed")
    if field == "title":
        return _gated_text(payload, "title", "title_completed")
    if field == "diff":
        return _gated_text(payload, "diff", "diff_completed")
    if field == "issue_body":
        return _gated_text(payload, "issue_body", "issue_body_completed")
    if field == "metadata":
        # The ``completion`` sub-object surfaces the per-section
        # ``*_completed`` flags so an operator inspecting via
        # ``--field metadata`` can see ALL fetcher state in one
        # query without having to iterate every ``--field``. The
        # worker's per-field reads still rely on the individual
        # field projections — this is for human/operator triage.
        #
        # Enumerates payload keys (rather than hardcoding the
        # name list) so a new ``*_completed`` flag added to the
        # writer's ``COMPLETION_FLAG_NAMES`` tuple appears here
        # automatically — keeps this reader script standalone
        # (no dispatcher / prefetch imports) while still tracking
        # writer-side additions without code changes. Stable
        # sort order so two consecutive ``--field metadata``
        # reads produce byte-identical output (worker / operator
        # diff tooling depends on this).
        completion = {
            k.removesuffix("_completed"): payload.get(k, True)
            for k in sorted(payload)
            if k.endswith("_completed")
        }
        return {
            "schema_version": payload.get("schema_version"),
            "pr_number": payload.get("pr_number"),
            "work_type": payload.get("work_type"),
            "work_group": payload.get("work_group"),
            "head_sha": payload.get("head_sha"),
            "head_ref": payload.get("head_ref"),
            "base_ref": payload.get("base_ref"),
            "listing_title": payload.get("listing_title"),
            "data_complete": payload.get("data_complete"),
            "completion": completion,
            "error_kinds": payload.get("error_kinds") or [],
            "created_at": payload.get("created_at"),
            "dispatcher_pid": payload.get("dispatcher_pid"),
        }
    if field == "ci":
        # ─── Contract divergence (intentional) ────────────────
        # Every OTHER field gated on a completion flag uses
        # _gated_list / _gated_text and returns ``_MISSING``
        # when the flag is False (empty stdout → worker falls
        # through to legacy GET). ``ci`` is the exception: it
        # ALWAYS emits the bundle when the sentinel exists at
        # all, and the worker is expected to inspect the inner
        # ``status_completed`` / ``detail_completed`` booleans
        # to decide what to do.
        #
        # Why diverge:
        # 1. CI carries TWO independent completion signals
        #    (status + detail). Collapsing them into one
        #    ``_MISSING`` would lose the case where the status
        #    summary fetch succeeded but the per-check detail
        #    fetch failed — the worker still has useful
        #    aggregate state ("PR is red"), it just lacks
        #    per-check evidence.
        # 2. A green PR legitimately has ``ci_detail=[]`` (no
        #    failing checks to enumerate) with
        #    ``ci_detail_completed=True``. Emitting that as a
        #    bundle is the right answer.
        #
        # Workers reading ``--field ci`` MUST check both
        # completion booleans before treating the bundle as
        # authoritative. See SKILL.md "ci field exception" for
        # the worker-side procedure.
        return {
            "status": payload.get("ci_status"),
            "checks": payload.get("ci_detail") or [],
            "status_completed": payload.get("ci_status_completed", True),
            "detail_completed": payload.get("ci_detail_completed", True),
        }
    if field == "comments":
        # Dispatch on ``work_type``: pr_fix / request_changes_pr
        # cycles populate ``pr_comments``; new_issue cycles populate
        # ``issue_comments``. Mixing them was a latent bug — if a
        # future change ever populated both (e.g. an issue_impl
        # cycle that also happened to have a stale PR), the previous
        # ``or`` chain would return the wrong one.
        if payload.get("work_type") == "issue_impl":
            return _gated_list(
                payload, "issue_comments", "issue_comments_completed"
            )
        return _gated_list(payload, "pr_comments", "pr_comments_completed")
    if field == "comments_digest":
        # The deterministic attempt-history digest the dispatcher
        # computed from the FULL comment list at prefetch time (see
        # _attempt_history.summarize_attempt_history). The verbatim
        # ``comments`` field only carries the bounded view; this is the
        # rollup of everything dropped from that window. Always emit the
        # dict when the sentinel exists — an empty ``{}`` means
        # "computed, zero attempts parsed", still authoritative. There
        # is no legacy GET to fall through to.
        return payload.get("pr_comments_digest") or {}
    if field == "reviews":
        return _gated_list(
            payload,
            "request_changes_reviews",
            "request_changes_reviews_completed",
        )
    if field == "comment_reviews":
        # R3.4 (2026-05-17): non-RC reviews (COMMENT/APPROVE) — the
        # reviewer's advisory feedback. Separate field from RC so
        # callers can keep the blocking-vs-advisory distinction.
        return _gated_list(
            payload,
            "comment_reviews",
            "comment_reviews_completed",
        )
    if field == "issues":
        return _gated_list(payload, "linked_issues", "linked_issues_completed")
    if field == "epic":
        # ``epic_completed=False`` means the fetch failed; fall
        # through. ``epic_completed=True`` with ``epic=None`` means
        # the dispatcher confirmed the PR has no Epic reference —
        # emit ``null`` so the worker skips its legacy GET.
        if not payload.get("epic_completed", True):
            return _MISSING
        return payload.get("epic")
    if field in (
        "compliance_gaps", "gate_preflight", "recent_implementer_push",
    ):
        # Deterministic / cross-cycle sections — BINARY contract
        # (not the generic three-case prefetch contract). The
        # dispatcher either computed this cycle (``*_completed=True``
        # AND a populated dict, emitted as JSON) or it did not
        # (both keys absent, emitted as empty stdout / worker falls
        # through). The detectors always produce a structured dict
        # when they run, so there is no "tried-but-empty" middle
        # state to surface.
        completed_key = f"{field}_completed"
        if not payload.get(completed_key):
            return _MISSING
        return payload.get(field)
    raise ValueError(f"unknown --field: {field}")


_ALLOWED_FIELDS = (
    "all",
    "description",
    "title",
    "diff",
    "metadata",
    "ci",
    "comments",
    "comments_digest",
    "reviews",
    "issues",
    "epic",
    "issue_body",
    "compliance_gaps",
    "gate_preflight",
    "recent_implementer_push",
)


def _emit(value: Any, field: str) -> None:
    """Render ``value`` to stdout in the format the worker expects.

    Three cases (see :data:`_MISSING` docstring for full rationale):

    - ``value is _MISSING`` → emit nothing. The worker reads empty
      stdout and falls through to its legacy GET.
    - ``value is None`` → emit ``null\\n``. The worker reads
      non-empty stdout and skips its legacy GET (the dispatcher
      authoritatively confirmed absence).
    - any other value → emit verbatim. Strings print as plain text
      with a trailing newline; everything else prints as
      pretty-printed JSON.
    """
    if value is _MISSING:
        return
    if value is None:
        # Authoritative "no data" answer (e.g. dispatcher confirmed
        # no Epic). The worker contract reads this as "skip the
        # legacy GET; there's nothing to fetch." JSON ``null`` is
        # parseable in any language the worker might use.
        sys.stdout.write("null\n")
        return
    if isinstance(value, str):
        # Plain-text contract for these fields so the worker can
        # pipe them straight into a file or display them inline
        # without an extra json parse.
        sys.stdout.write(value)
        # Ensure a trailing newline so terminal usage looks right.
        if not value.endswith("\n"):
            sys.stdout.write("\n")
        return
    # Structured fields: pretty-print JSON so the worker (or an
    # operator inspecting from the shell) gets a readable blob.
    json.dump(value, sys.stdout, indent=2, default=str)
    sys.stdout.write("\n")


def _cmd_read(args: argparse.Namespace) -> int:
    """Implement ``read``. Always exits 0 on the "no data" path —
    that's a normal condition (e.g. epic-field on a PR with no
    Epic reference). Exit 3 only on schema mismatch, which is a
    config issue the operator must notice."""
    payload, error = _load(int(args.pr))
    if error:
        print(f"warning: {error}", file=sys.stderr)
        if "schema_version" in error:
            return 3
        return 0
    if payload is None:
        # No sentinel — the dispatcher didn't prefetch (or cleaned
        # up already). Empty stdout is the "no data" signal.
        return 0
    try:
        value = _project_field(payload, args.field)
    except ValueError as e:
        print(f"error: {e}", file=sys.stderr)
        return 2
    _emit(value, args.field)
    return 0


def main(argv: list[str] | None = None) -> int:
    parser = argparse.ArgumentParser(
        prog="implementer_pr_context",
        description=(
            "Worker-side reader for the dispatcher's PR-context "
            "handoff sentinel. See module docstring for the full "
            "contract."
        ),
    )
    sub = parser.add_subparsers(dest="cmd", required=True)

    p_read = sub.add_parser(
        "read",
        help=(
            "Read one field from the PR's context handoff. Prints "
            "the field on stdout (plain text for description/title/"
            "diff/issue_body, JSON for everything else). Empty "
            "stdout means 'no data available — fall back to curl'."
        ),
    )
    p_read.add_argument(
        "--pr", required=True, type=int, help="PR number"
    )
    p_read.add_argument(
        "--field",
        required=True,
        choices=_ALLOWED_FIELDS,
        help="Which slice of the prefetched data to print",
    )

    args = parser.parse_args(argv)
    if args.cmd == "read":
        return _cmd_read(args)
    parser.error(f"unknown subcommand {args.cmd!r}")
    return 2  # pragma: no cover — argparse exits before reaching here


if __name__ == "__main__":
    sys.exit(main())