Repository navigation
Releases
ha 0.9.16
ha 0.9.16
Memory
- Deep Dream no longer deletes a stored procedure, runbook or set of steps: a long entry merged into nothing is demoted to a knowledge page word for word, or the delete is refused (#34).
Models
- Claude Opus, Sonnet and Haiku 5.5, Fable 5.2 and Mythos 5 / 5.1 / 5.2 on the Anthropic provider, with prices and the upstream catalog (#36).
Plugins
/plugins connect <id>connects API-key catalog services (GitHub, PagerDuty, Sonatype Guide, AWS DevOps Agent, Zoom): a masked prompt, the key kept in the credential file and sent as a bearer token, never written tomcp-servers.json(#35).
Fixes from bug report 20261120 (#38)
- A turn in one pane no longer writes the session log of a conversation that is still open in another pane.
/bug-reportlists a worker that is still running as running, not as a process without an exit record.--task/--sessionreports keep only that conversation's run steps.- The written-log notice uses one path separator and says when the log was another conversation's.
TUI (#39)
/resume+ Enter opens the session picker again;/newshows its flags first./heartbeat everyand/schedule addwait for the instruction instead of running with an empty one./refineshows its flags right after the name;/learn+ Enter runs.- Typing after
/featuresor/helpcloses the panel and brings back the cursor and suggestions. - MCP sign-in asks for the server's own scopes (Vercel
invalid_scope).
Verified
Gate: fmt, clippy, 2689 tests passed. Live on this ha.exe (openai-codex/gpt-6-luna, thinking max): diag 46/0, session-log 48/0, session-continuity 17/0, dream 22/0, claude-models 29/0, reread 16/0, commands 10/0, ui-stream 14/0, tui 15/0. Control on 0.9.15: diag 41/5 and session-log 37/11, the new checks among the failures.
ha 0.9.15
A review is told what was refused before and how each skill revision went, an empty step goes on, shell calls keep their place, and past commits are searchable.
Self-improvement (GUARD arXiv 2611.23476, SIM arXiv 2611.18856)
- Memory and prompt edits the host refused are kept per scope (at most 12) and every planner is told them in
<refused_before>. An edit refused for safety (a secret, steering text, a page's words made a rule) keeps only its reason, never its words. - A skill revision carries its
causeand what it should change (expect). Accepted revisions are kept as well as refused and rolled-back ones, and the planner reads each skill's history: cause, expected, observed. - Each learned skill is marked
regressed,unsolvedormastered. A regressed skill is narrowed with a condition rather than rewritten; a mastered one is only added to (a prune may still shorten it). - Not taken: GUARD's gate that tests a revision before it lands (the online trial with rollback stands for it).
Runtime (Harnesses That Improve Themselves, arXiv 2611.22299)
- A step with no text and no tool call is continued once, with the end of its reasoning; it used to end the turn with an empty answer.
run_shellandrun_processstart where the last shell call of the conversation left off, inside the workspace, with the PATH it set. The place is recorded before the command runs, soexit Nand an unclosed heredoc keep it; a native program's exit code is kept in PowerShell; one call of a conversation runs at a time; the shell starts by its full path; guards and Jev judge from the directory the command runs in; the approval shows and binds the directory and PATH.HA_SHELL_KEEP_LOCATION=offrestores the old behaviour.timeout_msof these tools is optional (60 s), as in 0.9.11.
Git history (RepoMem arXiv 2510.01003)
git_searchranks the repository's commits by message and changed files (BM25;parseHeader,parse_headerandparseheaderfind each other).git_showreads one commit, paged. Both are read-only.
Fixes
- A call waiting for its conversation's turn is reported queued and withdrawn when canceled.
- Windows guards: an absolute path to
.harnessin the\\?\form is refused.
Known
dream.sh: Deep Dream deletes a long procedure entry without evidence instead of demoting it to a knowledge page. It fails the same on 0.9.14 (control run); to be fixed in 0.9.16.
Verified
Gate (fmt, clippy, workspace tests) clean on the release tree. Live on this ha.exe with openai-codex/gpt-6-luna at max: jev.sh 33/0, memory.sh 44/0, learning.sh 49/0, guard.sh 11/0, memory-queue.sh 37/0, runtime-gaps.sh 10/0, git-history.sh 7/0, guards.sh 28/0, queue.sh 20/0, recall.sh 32/0, codemode.sh 11/0, bank.sh 18/0, self-eval.sh 18/0, step-judge.sh 10/0, memory-safety.sh 22/0, continual.sh 32/0, skill-trial.sh 13/0, dream.sh 19/3 (known, above). Control on 0.9.14: guard.sh 1/3, runtime-gaps.sh 5/5, git-history.sh 0/2 (no such tools), dream.sh 12/10.
ha 0.9.14
ha 0.9.14
Fixes from bug report 2026-11-19 11:52:
- A strategist's alternative can be taken. A
task_stateswitch that names only one of the strategist's alternatives takes it, as the list tells the model it may. - A refused
task_stateupdate is a plain failure, not "side effect may be uncertain; do not rerun automatically". Only a failed write stays an error. - No false plateau while children work. A goal's turn that ends while its children are still working is not evaluated. Before this, their reports' turns read as no progress and brought the strategist in seven times.
- Images are budgeted by their pixels, not their base64. A 153 KiB PNG had been estimated 58k tokens over, and an image over about 1 MB was refused as mandatory overflow. An image carried by a later message is counted too.
- An image named at the end of a sentence is attached whole.
Look at shot.png. …had sent the PNG cut off at 256 KiB, and the provider could not parse it. The full stop is not part of the path, an image found by its bytes is read whole up to the image ceiling, and that read is bounded. - The "save before the summary" ask comes once per approach. It re-arms only once the context is clear (60% or under) or summarized.
- The worker and the supervisor record their exit before they stop.
- Records name their conversation:
call.settledand regulation records. list_filesacceptspreview 0and a file's path.- The bug report says whether a claimed sleep job's process is still at it.
New in the app (PR #31):
- Every slash command's arguments complete, word by word.
/package install|remove|update|listworks in the chat. Its changes run on their own thread, and the menu offers local packages relative to the project.
Gate: 2568 passed, 0 failed. Live on this ha.exe (openai-codex/gpt-6-luna, thinking max): image 6/6, diag 40/40, children 15/15, search 26/26, strategy 32/32, compaction 25/25 (rerun; the first run's model wrote too little for a third summary), commands 10/10, tui 15/15. Controls on 0.9.13: image 2/6, diag 37/40, children 11/15, search 25/26.
ha 0.9.13
ha 0.9.13
Fixes from bug report 2026-11-18 16:08 and CI:
- One agent's process does not wait for another agent's. A parent's
run_shell/run_processno longer queues behind its child's. - A restart is named. The bug report says a process whose log starts before the computer last started ended with the computer, with the boot time, instead of listing it as still running or crashed. (The CLI denies
unsafe_codeinstead of forbidding it, for the oneGetTickCount64call; every library crate still forbids it.) - The person's language. The language of the conversation is the person's own messages', and the task state and the strategist's alternatives are written in it.
- The skill usage file loses no update on a slow machine: it is written under its lock (made in time, waited for up to a minute), opening the lock and the rename over the file retry a transient Windows refusal, and an unlocked or failed write is reported.
- CI: the process-tree test gives a loaded runner time to start its grandchild.
Gate: 2531 passed, 0 failed. Live on this ha.exe (openai-codex/gpt-6-luna, thinking max): reread 16/16, skill-trial 13/13, step-judge 10/10, queue 20/20, diag 38/38, strategy 32/32, children 14/14, goal 6/6.
Released before the last CI jobs finished, at the maintainer's request; their results are followed up.
ha 0.9.12
ha 0.9.12
From four papers (docs/architecture/COST_AND_SKILLS.en.md, with the gap tables):
- Re-reads not sent twice (arXiv 2611.12122). Within a turn, a
read_fileof a file the model still holds unchanged answers with a pointer to that result (fresh: truereads it again); a re-read after a small change sends only the changed lines; runs of 4+ identical lines of command output are shown once with their count.HA_REREAD=off;ha exec --jsonreportsreread. - Cost by part (arXiv 2611.12122).
ha exec --jsongivescost_usdsplit into cache read, cache write, uncached input and output for the run, its children and the memory passes, each call at its own model's price;/costand the bug report show the split. - Learned skills revised like weights (SkillOpt, arXiv 2605.23904). A revision is an
editof at most 4 structured edits; a whole rewrite of a proven skill is refused. The revision is on trial with the incumbent kept: two judged uses better accept it, two worse roll it back by itself (files, absorbed siblings and counters); a fix that replaces a failing revision keeps the sound incumbent. Refused, rolled-back and superseded revisions are shown to the refiner.ha learning reportcounts them. - What a skill costs, and its prune (EfficientSkill, arXiv 2611.19823). Steps, tool calls and output tokens per skill in play, against a baseline;
ha learning prune [--skill] [--model] [--thinking](and the curator's cadence,HA_SKILL_PRUNE=off) shortens proven skills that are long or costly, on trial and rolled back at more than 1.10x the incumbent's cost. - Each step judged in hindsight (arXiv 2611.17167). After a turn of 6+ tool calls, the helper model scores every action with the whole turn in view: did it help achieve the request? Kept in
step-judgments.jsonl, given to the next review as data, shown per learned skill and inha learning report.HA_STEP_JUDGE=off.
Fixes:
ha exec: a learned skill the model activated is in play for the review (the review job replaced the run's list with an empty one), and a skill on trial in play is always judged.- Goal: only the person answers the agent's question, and the host's own children do not wake the conversation.
- A message the user sends with
ha sendanswers the agent's open question (it was queued behind a run that was not running, andha send --waithung).
Live (openai-codex/gpt-6-luna, thinking max): reread.sh 16/16, skill-trial.sh 13/13, step-judge.sh 10/10, strategy.sh 31/31, queue.sh 19/19, children.sh 14/14, goal.sh 6/6. Control on 0.9.11: reread 8 of 16 fail, skill-trial 9 of 12 fail, step-judge 6 of 10 fail.
ha 0.9.11
Codemode: the model can write one program that calls its tools, and only what the program prints reaches the conversation. Codemode and the action bank are on by default.
Codemode (new, on by default)
- After pi's codemode, without its VM: a JavaScript interpreter written in Rust (the new
harness-codemodecrate) runs the program on a thread of its own. The program calls the turn's tools, the deferred ones included, asawait tools.<name>(args);Promise.allruns calls side by side. Onlytext()/console.logoutput and the returned value come back, so large intermediate results never enter the context. - Every call the program makes is checked against its tool's schema and crosses the same gate as the model's own calls: policy, command guard, hooks, approval, intent and receipt, and the learned rules. A refused call is an error the program can catch. Each call is a card of its own and one of the turn's executions.
- The program has nothing but its tools: no timers, network, files, processes or modules.
store()/load()keep small JSON values for the next program of the conversation, also afterha exec --resume. A first line// @options: {"timeout_ms", "max_output_tokens"}sets its limits (5 minutes and 60,000 characters by default). /codemode on|off|status, settings.jsoncodemode,HA_CODEMODE. See docs/architecture/CODEMODE.en.md, which maps pi's package to ha, and the operator guide's 12.27.
Action bank
- On by default.
/bank off, settings.json"actionBank": falseorHA_ACTION_BANK=offturns it off;/bank statussays whether the setting or the default decided.
Strategy
- A strategist that sends its parent the diagnosis and alternatives with
agent_messageand ends with a summary is still heard: the consult reads the object from its message. On the release candidate the strategist did so, the plateau got no alternatives and the agent never switched.
Tools and continuity
run_shellandrun_processrun with a 60 s timeout when a call names none, instead of refusing the call;run_process's args may be left out.- A worker that stops lets go of the conversations it held and the logs it owed, so a conversation started right after is shown the task the last one left unfinished.
Checked
- Gate: 2427 passed, 0 failed. Live on
openai-codex/gpt-6-luna(thinking max) on thisha.exe: codemode 12/0 (0.9.10 as control: no codemode), strategy 25/0, bank 16/0, memory 44/0, continual 32/0; on the candidate before the strategist fix: guards 28/0, skill-activate 11/0, recall 32/0, related-logs 33/0, session-log 42/0, learning 49/0.
ha 0.9.9
A new conversation is shown the earlier session logs that name what it asks, logs are revised by what later conversations found, and a learned skill is not offered where words cannot tell.
Session logs
After ReHarness (arXiv 2610.12668).
- Forward read: a new conversation's first turn carries up to 3 earlier session logs that name what its request names (
related-session-logs, in the app and inha exec, which showed no log at all before). A word of the request tells logs apart when at most a quarter of them hold it; a log is related when its topic, keywords or projects hold one; logs are ranked by the rarest such word each holds, so words many logs share cannot lift the log of another problem over the logs of the one named. No word list; with fewer than four logs none is chosen by words. - Backward revision: the log a conversation writes says what held of the logs it was given, and each such line is appended to that log under
## Later Conversations, dated and naming the conversation. The line names what held instead when the finding did not. - The log states the finding itself (the rule, the command, the value), never a pointer to the conversation; puts the problem's own name first in its keywords; and, for a problem that may come back, records what carries over apart from that session's values. A finding used from a given log is restated as it now stands.
ha execwaits up to 90 seconds at exit for the session log (30 before); a log that takes longer stays owed to the next start.
Learned skills
- A skill is recalled whole only in a library of four or more: with fewer, every word seemed to tell.
- The catalog leaves out a skill the conversation already holds whole, and
activate_skillon such a skill loads nothing again and says so (one benchmark run made 110 such calls; the same run now makes 6). - A review does not patch a skill with a lesson about a different problem; a skill offered where it does not fit has its description narrowed.
Tests
- New live suite
related-logs.sh.
Verified
Gate (fmt, clippy, workspace tests) clean on 166d591's tree. Live on this ha.exe with openai-codex/gpt-6-luna at max: related-logs.sh 33/0, session-log.sh 42/0, recall.sh 32/0, continual.sh 32/0, strategy.sh 25/0, learning.sh 49/0, memory.sh 44/0. Control on 0.9.8: related-logs.sh 8/12.
Continual-ARC (6 tasks x 4 visits, thinking medium), solved / budget share: 0.9.7 0.58 / 0.583, 0.9.8 0.46 / 0.685, a build of this branch before the last two fixes 0.50 / 0.604. Within noise at this size, and no run shows later visits getting cheaper.
ha 0.9.10
A skill the model activates now reaches it, memory no longer hides rows of digits as card numbers, and two jobs written in one millisecond are merged in the right order.
Learned skills
- The result of
activate_skillholds the skill's text, in the<skill name="…">tag a recalled skill is given in. Since 0.9.7 a turn sends the packet of its first step at every step, so a skill activated mid-turn was added to a packet that was not sent while the result said its instructions were in the conversation. In one recordedha execconversation of 32 requests the model calledactivate_skillfor one skill more than twenty times and its text was in none of them. - A skill whose activation result the conversation still sends whole counts as held, in the app and after
ha exec --resume: the catalog leaves it out andactivate_skillanswers that it is already there. This holds when the library has no learned skill too. - A later turn's packet leaves out the block of a skill held this way, which was otherwise given twice; a skill the user ran with
/skill:keeps its block. - A skill longer than one tool result is cut inside its tag and says so.
Memory
- Redaction: a card number is written whole or in groups of three digits or more. A row of 13-16 single digits (a grid row) passed the Luhn check and was replaced, and session logs then recorded that rows of the input were redacted.
Background queue
- A failed job joins the job of its key that came in meanwhile, whatever the clock says. The newer of two jobs was decided by a time in milliseconds; written within one millisecond, the failed job's span replaced the newer one's (the ubuntu CI job failed on it).
Tests
- New live suite
skill-activate.sh.
Verified
Gate (fmt, clippy, workspace tests) clean. Live on this ha.exe with openai-codex/gpt-6-luna at max: skill-activate.sh 11/0, recall.sh 32/0, related-logs.sh 33/0, session-log.sh 42/0, continual.sh 32/0, strategy.sh 25/0, learning.sh 49/0, memory.sh 44/0. Control on 0.9.9: skill-activate.sh 8/3.
Continual-ARC (6 tasks x 4 visits, thinking medium), solved / budget share: 0.9.7 0.58 / 0.583, 0.9.8 0.46 / 0.685, 0.9.9 0.50 / 0.654, this code 0.58 / 0.557, with the fourth visit the cheapest (0.500). Paired with 0.9.9: budget share -0.096 (s.e. 0.068). Six instances per visit: indicative only.
ha 0.9.8
Memory that keeps what a conversation found (SoMem, Continual-ARC), learned skills that are recalled for a request, and reviews that keep more of what they propose.
Memory and knowledge
After SoMem (arXiv 2611.11929) and Continual-ARC (arXiv 2611.14505).
- A knowledge page's claims carry footnote evidence and anchors; the wiki has a timeline and a recent view.
- What survives a conversation (P1-P6): a script a skill carries is recorded as tested with the SHA-256 of the file that ran, and a task that may come back is saved before the conversation ends.
- P7: the review gate's note reaches the planner as a note, not as the user's request. The planner quoted the note as evidence and the host refused the edits (39 of about 50 in one benchmark run).
- P8: what solved a problem that may come back is kept under the problem's own name. The
memorytool tells the agent to title its own write by that name, and a review retitles an entry that holds the solution without it. - An upkeep model call (review, gate, evolution, wiki, lessons) cut short by a transient provider error is made once more, after the provider's wait (at most 30 s) or 2 s; the bug report lists it as
refinecall_retried.
Learned skills
- A skill is recalled for the goal's objective and what the user typed, not for text the host or a child wrote.
- A skill is put whole once per conversation; later turns carry one line. It is put whole only when the request holds words of its name or description.
- When no skill stands out, the turn carries the catalog of learned skills once.
- The user sees
learned skill X offered for this request;offeredanduseare counted apart. ha execrecalls skills and hasactivate_skill,list_skillsandread_skill_file.
Goals
- A strategist that names nothing, or fails, is asked once per plateau, not at every evaluation.
Tests
- New live suites
recall.sh,continual.sh,evidence.sh. - The knowledge catalog test no longer depends on how fast it writes (the ubuntu CI job).
Verified
Gate (fmt, clippy, workspace tests) clean on c07b35c. Live on this ha.exe with openai-codex/gpt-6-luna at max: recall.sh 32/0, knowledge.sh 48/0, learning.sh 49/0, continual.sh 32/0, memory.sh 44/0, evidence.sh 29/0, wiki.sh 23/0, strategy.sh 25/0, bootstrap.sh 12/0, memory-safety.sh 22/0, goal.sh 6/0, children.sh 14/0. Control on 0.9.7: recall.sh 18/14, evidence.sh 10/19, continual.sh 18/14, goal.sh 5/1.
ha 0.9.7
Fixes from a bug report of a long goal session, a bug report that is easier to trace, and a prompt cache that is not rewritten in the middle of a turn.
From bug report 20261112
- A summary the user stopped with Ctrl-C is recorded as canceled (
compaction.earlier_summary_canceled,turn_summary_canceled), not as an error, and raises no notice. - A memory review is told what a
[Host]:block is: a turn the host started, never the user's words. Reviews quoted the goal's continuation as a correction's evidence and had every edit refused. - A checkpoint of the whole conversation starts from the summary the turns kept and summarises only what came after it. It wrote the conversation again from its first message, four minutes before a turn.
- A request the provider refused as over its window is tried again on a run that is still open. The run read
failedwhile the second try worked. - A background command's end says which command ended and with what exit code.
- The eight fixes of the first pass over the same report (commit 70abf55).
The bug report
review.appliedcarries why a review's edits were refused, and SUMMARY counts the reasons.- SUMMARY names what the host sent into running turns by kind, the checkpoints written before a turn and how long they took, the summaries the user canceled, and a run or a command a step left open.
- It counts the requests that rewrote what was already sent, and the tool results dropped.
Prompt cache
After AKV (arXiv 2610.28577, thu-nics/AKV). AKV drops and repositions KV states inside the serving engine; a hosted provider caches a request by its start, so its rules are kept in the text ha sends.
- A turn of a compacted conversation sends its first packet at every step. Every step built it again.
- A tool result sent shortened is sent as that same text by every later step and every later turn, so it is rewritten once.
- A request that rewrote what the one before sent is recorded as
cache.prefix_rewritten, with what it rewrote. HA_CONTEXT_DROP=keep[:n]keeps a turn's newest n tool results (12 by default) and drops older ones a batch at a time.toolgives the modeldrop_context. Both are off by default: a drop costs one re-read of the cache. Reasoning and the model's own messages are never dropped.- Each model call records the tokens written to the cache.
- Not possible on a hosted API, so not done: KV-level drop and reposition, the radix tree, drop-aware eviction.
Other
- The fixture host gets
ha's 8 MiB main stack; a debug build's turn overflowed the default one. - New live suite
scripts/testing/live-tests/context-drop.sh.
Verified
Gate (fmt, clippy, workspace tests) clean. Live on this ha.exe with openai-codex/gpt-6-luna at max: context-drop.sh 15/0, prime.sh 17/0, diag.sh 37/0, queue.sh 19/0, memory.sh 44/0, compaction.sh 25/0. Control: context-drop.sh on 0.9.6 is 4/11.