Fix pier verify silently dropping trajectory data

[jbragg]

Problem

When pier verify couldn't determine which agent was used — because
none was registered, or --agent wasn't passed — it silently saved
the reward without the trajectory (conversation, tool use, cost data).
Users would see the score and assume everything was captured.

Changes

Fix trajectory loss: pier verify auto-infers the agent from the
session (set by pier start --agent) and prints which one it's using.
No agent → reward still saved, but with clear guidance on how to
capture the trajectory. Multiple agents → error asking to pick one.

Remove agent-specific session detection from pier: Container mode
now delegates to Harbor's populate_context_post_run directly
(extract_agent_context), removing find_container_agent_session_dir
and its per-agent detection code. extract_agent_logs calls
extract_agent_context after bridging.

Container --session-dir: Now refers to a path inside the
container (copied via docker cp). Supports agents not installed
through pier, e.g. baked into the image.

--session-dir requires -a: When --session-dir is explicitly
passed, -a/--agent is required (session inference is skipped to avoid
mismatching the agent with unrelated session files). Agent
auto-detection from file patterns is also removed.

Messaging: pier start without --agent hints about agent
installation and trajectory capture. --agent in host mode errors
with explanation. Extraction failure in container mode warns explicitly.

--no-mount doc fix: Clarified that Harbor's internal mounts
(agent logs, verifier output) still write to the host under .pier/.

Architecture note in README: pier stays agent-agnostic — agent
details belong in Harbor.

Breaking changes

• pier capture/verify --session-dir without -a now errors
• In container mode, --session-dir is now interpreted as a container
  path (previously a host path)

[Copilot]

Pull request overview

Fixes pier verify dropping trajectory data when an agent isn’t explicitly selected by inferring the agent from the pier session (when unambiguous) and delegating container-mode log extraction to Harbor. Updates CLI behavior/messaging and documentation to make trajectory capture expectations explicit, and aligns the test suite with the new flow.

Changes:

• Infer agent for pier verify/pier capture from the session when exactly one is registered; error on multiple; provide clearer messaging when none.
• Remove pier-side agent/session-dir detection heuristics; introduce extract_agent_context() and route container-mode extraction through Harbor directly.
• Update README and tests for new --session-dir semantics and required flags.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
README.md	Documents new trajectory capture behavior, --session-dir semantics, and pier’s agent-agnostic architecture stance.
pier/cli.py	Implements agent inference, container-path --session-dir copying via docker cp, and revised capture/verify assembly logic.
pier/harbor_bridge.py	Removes agent-detection helpers and adds extract_agent_context() to delegate extraction to Harbor.
pier/tests/test_pier_cli.py	Adds/updates tests for agent inference, multi-agent errors, container --session-dir copy, and -a requirements.
pier/tests/test_harbor_bridge.py	Updates default agent name expectation and removes tests for deleted detection helpers.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

This PR fixes pier verify silently dropping trajectory data when the agent can’t be determined, and moves agent/session detection responsibilities toward Harbor to keep pier agent-agnostic.

Changes:

• Infer the agent for pier verify/pier capture from the saved session (when unambiguous), emit clearer messaging, and error on multi-agent ambiguity.
• In container mode, delegate trajectory extraction to Harbor via extract_agent_context, and reinterpret --session-dir as a container path copied out via docker cp.
• Remove pier-side agent auto-detection heuristics and update docs/tests accordingly.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
README.md	Updates CLI docs/messaging around --no-mount, agent inference, and --session-dir semantics; adds architecture note.
pier/cli.py	Implements agent inference, new container session copy helper, revised verify/capture flow, and improved user-facing guidance.
pier/harbor_bridge.py	Removes agent/session-dir heuristics and introduces extract_agent_context to delegate parsing to Harbor.
pier/tests/test_pier_cli.py	Adds/updates coverage for agent inference, ambiguity errors, container --session-dir copy behavior, and -a requirement.
pier/tests/test_harbor_bridge.py	Updates trial-result default agent name expectation; removes tests for deleted heuristics.

Comments suppressed due to low confidence (1)

pier/harbor_bridge.py:909

• extract_agent_context() calls AgentName(agent_name) / AgentFactory.create_agent_from_name(...) outside the try block. If the agent name is invalid (user typo) or Harbor rejects the logs dir, this will raise and can crash pier verify/pier capture instead of returning None as the docstring promises. Wrap agent creation and name parsing in the same try/except (logging a warning and returning None) so extraction failures remain non-fatal and consistent.

    from harbor.agents.factory import AgentFactory
    from harbor.models.agent.context import AgentContext
    from harbor.models.agent.name import AgentName

    agent = AgentFactory.create_agent_from_name(
        AgentName(agent_name), logs_dir=logs_dir
    )
    context = AgentContext()

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[mikerjacobi]

Bug: _tar_copy_from_container swallows stderr on failure

Carried in from the merge of #2. In _tar_copy_from_container the docker exec process is started without stderr=subprocess.PIPE, so when the tar fails the error message is just:

tar copy from container failed (create=1, extract=0)

No docker error text, no indication of what actually went wrong (container stopped, path doesn't exist, permissions, etc.).

_tar_copy_to_container right next to it handles this correctly — it captures stderr and surfaces it in the ClickException. Same pattern should be applied here:

with subprocess.Popen(
    ["docker", "exec", container, "tar", "-cf", "-", ...],
    stdout=subprocess.PIPE,
    stderr=subprocess.PIPE,   # add this
) as tar_create:
    ...
tar_create_stderr = tar_create.stderr.read()
# then reference it in the error message like _tar_copy_to_container does