Make targets for individual Phase 1 subphases (phase-1a, phase-1b, phase-1c)

[pruiz]

Motivation

Currently make phase-1 runs all three subphases (1a → CodeQL → 1b → 1c) in a single orchestrated run. There is no way to run or re-run a single subphase independently. This is needed for:

• Recovery: If Phase 1b fails mid-turn, the user must re-run the full make phase-1 which re-gates 1a (already done) and re-runs the entire chain.
• Iteration: During development, re-running a single subphase avoids wasted work.
• Targeted fixes: After manual edits to a specific artifact, running just the relevant subphase for validation/repair is cleaner than the full chain.
• Debugging: Isolating subphase issues requires running them individually.

Design considerations

Runner changes (tools/codecome/phase_1.py)

run_phase_1() currently orchestrates all three subphases internally. It needs to expose individual entry points:

def run_phase_1a(args, console, rendering_ctx, runner, base_url) -> int:
    """Run Phase 1a only with gate before + after."""

def run_phase_1b(args, console, rendering_ctx, runner, base_url) -> int:
    """Run Phase 1b only (assumes 1a artifacts exist, gates that)."""

def run_phase_1c(args, console, rendering_ctx, runner, base_url) -> int:
    """Run Phase 1c only (assumes 1b artifacts exist, gates that)."""

Dispatch (tools/run-agent.py or tools/codecome/harness.py)

--phase 1a / --phase 1b / --phase 1c need to map to the right runner functions. Currently run_phase_mode in harness.py only handles --phase 1 → run_phase_1(). Needs to accept subphase dispatch.

Makefile targets

phase-1a: env-check
    @$(PYTHON) tools/gate-check.py 1a
    @$(PYTHON) tools/run-agent.py --phase 1a --label "Phase 1a: Target Profile" --agent recon

phase-1b: env-check
    @$(PYTHON) tools/gate-check.py 1b
    @$(PYTHON) tools/run-agent.py --phase 1b --label "Phase 1b: Detailed Reconnaissance" --agent recon --prompt-file prompts/phase-1b-recon.md

phase-1c: env-check
    @$(PYTHON) tools/gate-check.py 1c
    @$(PYTHON) tools/run-agent.py --phase 1c --label "Phase 1c: Sandbox Bootstrap" --agent recon --prompt-file prompts/phase-1c-sandbox.md

Gate changes

tools/gate-check.py already supports 1a, 1b, 1c via run_from_cli() → check_phase_1a() etc. No gate changes needed.

Dependency awareness

• make phase-1b must gate on Phase 1a artifacts (target-profile.md, build-model.md) — already handled by check_phase_1b via REQUIRED_NOTES_1B
• make phase-1c must gate on Phase 1b artifacts — already handled by check_phase_1c
• No automatic chaining (e.g., make phase-1b does not auto-run phase-1a)

Server lifecycle

Currently run_phase_1() starts the opencode server once and reuses it across all three subphases. Individual subphase runners would each start/stop their own server. This is acceptable — the overhead of starting the server is negligible compared to model response time.

PROMPT_EXTRA / PROMPT_EXTRA_FILE

These must be passed through to subphase runners, same as they are for the full make phase-1.

Acceptance criteria

• make phase-1a runs only Phase 1a with gate before and after
• make phase-1b runs only Phase 1b with gate before and after
• make phase-1c runs only Phase 1c with gate before and after
• make phase-1 continues to work (runs all three in sequence)
• CODECOME_THINKING=1 PROMPT_EXTRA="..." work on all three targets
• Existing tests pass; new tests cover subphase dispatch
• make help lists the new targets
• Phase 1a/1b/1c gate labels match the subphase names